[ { "id": "qwen3-8-27b", "name": "Qwen3.8-27B", "maker": "Alibaba / Qwen", "categories": [ "multimodal", "text", "code", "vision" ], "parameters": "27B", "context": "262,144 native · up to 1M with YaRN", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A 27B dense vision-language model for reasoning, coding, document work and long-horizon agents. It accepts images and video and supports 262K native context, extendable to 1M with YaRN.", "architecture": "64-layer dense hybrid Transformer; Gated DeltaNet plus global attention and multi-token prediction", "input": "Text, image and video → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen3.8-27B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": 89.2, "lcb": 90.3, "swe": null } }, "benchmarkNotes": "SWE-bench Pro 61.7 · Terminal-Bench 2.1 73.0 · OSWorld-Verified 84.3 · AndroidWorld 81.9 · IFBench 79.5 · MathVision 90.0 · LiveCodeBench is v6 (publisher card, xhigh)", "aliases": [], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3.8-27B", "note": "" } ], "benchmarkHighlights": [ { "label": "SWE-bench Pro", "value": "61.7%" }, { "label": "Terminal-Bench 2.1 · Terminus", "value": "73.0%" }, { "label": "NL2Repo-Bench", "value": "42.3%" }, { "label": "QwenSWEBench · avg@3", "value": "79.0%" }, { "label": "CoWorkBench", "value": "70.7%" }, { "label": "IFBench", "value": "79.5%" }, { "label": "Humanity’s Last Exam", "value": "30.8%" }, { "label": "OSWorld-Verified", "value": "84.3%" }, { "label": "AndroidWorld", "value": "81.9%" }, { "label": "MathVision · without CI", "value": "90.0%" }, { "label": "OmniDocBench 1.5", "value": "91.1%" } ] }, { "id": "qwen3-coder-32b", "name": "Qwen3-Coder-30B-A3B-Instruct", "maker": "Alibaba", "categories": [ "code" ], "parameters": "30.5B total · 3.3B active", "context": "262,144 native · up to 1M with YaRN", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "An instruction-tuned coding model for repository understanding, code generation and tool-based engineering. The official checkpoint is Qwen3-Coder-30B-A3B-Instruct: 30.5B total parameters, 3.3B active, 256K native context, and up to 1M with YaRN.", "architecture": "Sparse MoE Transformer; 128 experts with 8 active", "input": "Text and code → text/code", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen3-Coder-32B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [ "Qwen3-Coder-32B" ], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct", "note": "" } ], "old": "Qwen3-Coder-32B" }, { "id": "kimi-k2-instruct", "name": "Kimi-K2-Instruct", "maker": "Moonshot AI", "categories": [ "code" ], "parameters": "1T (32B active)", "context": "128K", "license": "Modified MIT", "openness": "Open weights · permissive licence", "description": "Moonshot AI’s instruction-tuned sparse MoE for reasoning, coding and agent workflows. It contains 1T total parameters with 32B active per token, supports 128K context and tool calling.", "architecture": "Sparse MoE Transformer with 384 experts", "input": "Text → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Kimi-K2-Instruct" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 81.1, "gpqa": 75.1, "lcb": 53.7, "swe": 65.8 } }, "benchmarkNotes": "Publisher instruction-model table. GPQA-Diamond uses Avg@8; LiveCodeBench v6 is Pass@1; SWE-bench Verified is the single-attempt agentic result with no test-time compute. The separately reported 71.6% uses multiple attempts and an internal reranker.", "aliases": [], "provenance": [ "official model card" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct", "note": "" } ], "benchmarkHighlights": [ { "label": "AIME 2024 · Avg@64", "value": "69.6%" }, { "label": "AIME 2025 · Avg@64", "value": "49.5%" }, { "label": "MATH-500", "value": "97.4%" }, { "label": "τ²-Bench · retail · Avg@4", "value": "70.6%" }, { "label": "τ²-Bench · airline · Avg@4", "value": "56.5%" }, { "label": "τ²-Bench · telecom · Avg@4", "value": "65.8%" }, { "label": "SWE-bench Verified · multi-attempt reranked", "value": "71.6%" } ] }, { "id": "glm-5-2-osl", "name": "GLM-5.2", "maker": "Open-Source-Lab / Z.ai", "categories": [ "code", "experimental" ], "parameters": "744B total · 40B active", "context": "1M", "license": "MIT", "openness": "Hosted listing · base model and licence identified", "description": "Z.ai’s long-horizon reasoning and coding model for large repository work and agentic tool use. Its publisher specifies 1M context and MIT licensing; the OSL entry is a hosted route to GLM-5.2.", "architecture": "Sparse MoE Transformer with IndexShare attention and multi-token prediction", "input": "Text → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/zai-org/GLM-5.2" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/GLM-5.2-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/zai-org/GLM-5.2", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "deepseek-v4-osl", "name": "DeepSeek-V4-Flash", "maker": "Open-Source-Lab / DeepSeek", "categories": [ "text", "code" ], "parameters": "284B total · 13B active", "context": "1M tokens", "license": "MIT", "openness": "Open weights · permissive licence", "description": "DeepSeek-V4-Flash is a 284B-parameter Mixture-of-Experts model with 13B active parameters and a 1M-token context. DeepSeek positions Flash as the faster, lower-cost V4 option, with reasoning close to V4-Pro and similar performance on simple agent tasks.", "architecture": "Mixture-of-Experts Transformer with Compressed Sparse Attention (CSA), Heavily Compressed Attention (HCA) and mHC", "input": "Text → text; reasoning, coding and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/DeepSeek-V4-OSL" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 86.2, "gpqa": 88.1, "lcb": 91.6, "swe": 79 } }, "benchmarkNotes": "Publisher mode table: MMLU-Pro, GPQA Diamond and LiveCodeBench values are for V4-Flash Max reasoning mode. SWE-bench Verified is reported as resolved. Scores change with reasoning mode and evaluation protocol.", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash", "note": "Publisher specification: 284B total parameters, 13B activated, 1M context, and MIT license." } ], "benchmarkHighlights": [ { "label": "Humanity’s Last Exam · Max", "value": "34.8%" }, { "label": "Terminal Bench 2.0 · Max", "value": "56.9%" }, { "label": "SWE-bench Pro · Max", "value": "52.6%" }, { "label": "SWE-bench Multilingual · Max", "value": "73.3%" }, { "label": "BrowseComp · Max", "value": "73.2%" }, { "label": "Humanity’s Last Exam with tools · Max", "value": "45.1%" } ] }, { "id": "deepseek-v3-2-osl", "name": "DeepSeek-V3.2", "maker": "Open-Source-Lab / DeepSeek", "categories": [ "multimodal", "experimental" ], "parameters": "671B total · 37B active", "context": "128K", "license": "MIT", "openness": "Hosted listing · base model and licence identified", "description": "DeepSeek’s reasoning and agentic text model built around DeepSeek Sparse Attention. It targets coding, problem solving and tool use and is released under MIT.", "architecture": "Sparse MoE Transformer with DeepSeek Sparse Attention (DSA)", "input": "Text → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/DeepSeek-V3.2-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "deepseek-v4-pro-osl", "name": "DeepSeek-V4-Pro", "maker": "Open-Source-Lab / DeepSeek", "categories": [ "multimodal", "experimental" ], "parameters": "1.6T total · 49B active", "context": "1M", "license": "MIT", "openness": "Hosted listing · base model and licence identified", "description": "DeepSeek’s Pro model for long-context reasoning, coding and agent tasks. It has 1.6T total parameters, 49B activated parameters and 1M context; the -OSL Poe listing is the same base model.", "architecture": "Sparse MoE with CSA/HCA hybrid attention and mHC residual connections", "input": "Text → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/DeepSeek-V4-Pro-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [ "DeepSeek-V4-Pro-OSL" ], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro", "note": "Canonical name and specifications confirmed against DeepSeek’s publisher card; this is the same base model as the OSL listing." } ], "old": "DeepSeek-V4-Pro-OSL" }, { "id": "mistral-7b-osl", "name": "Mistral-7B-Instruct-v0.2", "maker": "Open-Source-Lab / Mistral", "categories": [ "text", "experimental" ], "parameters": "7B", "context": "32K", "license": "Apache 2.0", "openness": "Hosted listing · base model and licence identified", "description": "A 7B instruction-tuned language model for local chat, summarization and code assistance. This listing maps to Mistral-7B-Instruct-v0.2, with 32K context and Apache 2.0 licensing.", "architecture": "Decoder-only Transformer with grouped-query attention", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mistral-7B-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "mistral-nemo", "name": "Mistral-Nemo-Instruct-2407", "maker": "Open-Source-Lab / Mistral", "categories": [ "multimodal", "experimental" ], "parameters": "12B", "context": "128K", "license": "Apache 2.0", "openness": "Hosted listing · base model and licence identified", "description": "A 12B instruction model co-developed by Mistral AI and NVIDIA. It supports 128K context, multilingual text generation and function calling under Apache 2.0.", "architecture": "Decoder-only Transformer with multilingual tokenizer and function calling", "input": "Text → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mistral-NeMo" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "mistral-medium-3-5", "name": "Mistral Medium 3.5", "maker": "Open-Source-Lab / Mistral", "categories": [ "multimodal", "experimental" ], "parameters": "128B dense", "context": "256K", "license": "Modified MIT", "openness": "Hosted listing · base model and licence identified", "description": "Mistral’s 128B flagship for reasoning, coding and agent workflows. It accepts text and images, supports 256K context and function calling, and uses a modified MIT licence.", "architecture": "Dense Transformer with vision encoder and native function calling", "input": "Text and images → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mistral-Medium-3.5" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": 77.6 } }, "benchmarkNotes": "Publisher-reported: SWE-bench Verified 77.6% · τ³-Telecom 91.4%.", "aliases": [], "provenance": [ "existing catalogue", "publisher model card" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B", "note": "Publisher benchmark figures are reported on the linked card; a hosted service can use a different serving configuration." } ], "benchmarkHighlights": [ { "label": "τ³-Telecom", "value": "91.4%" } ] }, { "id": "gemma-4-26b", "name": "Gemma 4 26B A4B IT", "maker": "Open-Source-Lab / Google", "categories": [ "text", "experimental" ], "parameters": "25.2B total · 3.8B active", "context": "256K", "license": "Apache 2.0", "openness": "Hosted listing · base model and licence identified", "description": "An instruction-tuned Gemma 4 MoE for chat, reasoning, coding and image understanding. The 26B A4B variant has 25.2B total and 3.8B active parameters, 256K context, and text/image input; it does not support audio.", "architecture": "MoE with hybrid local sliding-window and global attention; vision encoder", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Gemma-4-31B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "nemotron-3-super", "name": "Nemotron 3 Super 120B A12B", "maker": "Open-Source-Lab / NVIDIA", "categories": [ "text", "experimental" ], "parameters": "120B total · 12B active", "context": "Up to 1M (256K default configuration)", "license": "NVIDIA Nemotron Open Model License", "openness": "Hosted listing · base model and licence identified", "description": "NVIDIA’s reasoning model for long-context and agentic workloads. It combines Mamba-2, latent MoE and attention layers, supports up to 1M context, and is released under the NVIDIA Nemotron Open Model License.", "architecture": "LatentMoE hybrid: Mamba-2, MoE, attention and multi-token prediction", "input": "Text → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Nemotron-3-Super" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "wan-2-2-osl", "name": "Wan 2.2", "maker": "Open-Source-Lab / Alibaba", "categories": [ "video", "multimodal", "experimental" ], "parameters": "Family: 5B dense TI2V or 27B total A14B MoE", "context": "Video generation · up to 720p; 5 seconds for T2V-A14B", "license": "Apache 2.0", "openness": "Hosted listing · base model and licence identified", "description": "Wan 2.2 is a family of video-generation models, not one checkpoint. It includes 14B-class MoE text-to-video and image-to-video variants and a 5B text-and-image-to-video model, with outputs up to 720p.", "architecture": "Video diffusion family; dual-expert MoE A14B and dense TI2V-5B variants", "input": "Text or image → video", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/WAN-2.2-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B", "note": "The OSL listing does not identify a specific Wan 2.2 variant; source describes the family and its variants." } ] }, { "id": "trellis-3d-osl", "name": "TRELLIS Image Large", "maker": "Open-Source-Lab / Microsoft", "categories": [ "3d", "experimental" ], "parameters": "2B", "context": "One image → 3D assets", "license": "MIT", "openness": "Hosted listing · base model and licence identified", "description": "Microsoft Research’s image-to-3D model turns an image into textured assets and multiple 3D representations. The published large checkpoint is about 2B parameters and is MIT licensed.", "architecture": "Sparse 3D flow-matching Transformer with structured latent representations", "input": "Image → 3D shape / texture assets", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/microsoft/TRELLIS-image-large" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Trellis-3D-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/microsoft/TRELLIS-image-large", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "llama-3-3-70b", "name": "Llama-3.3-70B-Instruct", "maker": "Meta", "categories": [ "text", "code" ], "parameters": "70B dense", "context": "128K", "license": "Llama 3.3 Community License", "openness": "Open weights · custom or non-commercial licence", "description": "Dense 70B instruction-tuned model with 128K context and eight officially supported languages. Widely hosted community workhorse for chat, tool calling and coding; released under the Llama 3.3 Community License (open weights, not OSI-approved).", "architecture": "Autoregressive dense Transformer with grouped-query attention", "input": "Text → text", "destinations": [ { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/llama-3.3-70b-versatile" }, { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/meta/llama-3_3-70b-instruct" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/meta-llama/llama-3.3-70b-instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Llama-3.3-70B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 68.9, "gpqa": 50.5, "lcb": null, "swe": null } }, "benchmarkNotes": "MMLU 86.0 · HumanEval 88.4 · MATH 77.0 · IFEval 92.1 · BFCL v2 77.3 (publisher card)", "aliases": [ "Llama 3.3 70B" ], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct", "note": "" } ], "benchmarkHighlights": [ { "label": "MMLU · CoT", "value": "86.0%" }, { "label": "HumanEval · pass@1", "value": "88.4%" }, { "label": "MATH · CoT", "value": "77.0%" }, { "label": "IFEval", "value": "92.1%" }, { "label": "BFCL v2", "value": "77.3%" } ] }, { "id": "llama-3-2-vision", "name": "Llama-3.2-11B-Vision-Instruct", "maker": "Meta", "categories": [ "vision" ], "parameters": "11B", "context": "128K", "license": "Llama 3.2 Community License", "openness": "Open weights · custom or non-commercial licence", "description": "Meta’s 11B vision-language assistant for image understanding and chat. It accepts text and images, produces text, supports 128K context, and uses Meta’s Llama 3.2 Community License.", "architecture": "Llama language backbone with vision adapter and cross-attention", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Llama-3.2-11B-Vision" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct", "note": "" } ] }, { "id": "qwen2-5-vl-72b", "name": "Qwen2.5-VL-72B", "maker": "Alibaba", "categories": [ "vision" ], "parameters": "72B", "context": "32,768 native · extendable with YaRN", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A 72B visual assistant for image/document understanding, OCR, charts, object localization and video question answering. The publisher sets 32K as the default context and documents YaRN extension for longer text.", "architecture": "Vision-language Transformer with dynamic-resolution ViT and temporal video encoding", "input": "Text, images and video → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen2.5-VL-72B-Instruct" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct", "note": "" } ] }, { "id": "mistral-small-4", "name": "Mistral Small 4", "maker": "Mistral AI", "categories": [ "text", "code", "vision", "multimodal" ], "parameters": "119B total · 6B active", "context": "262,144", "license": "Apache 2.0", "openness": "Open weights · Apache 2.0", "description": "A 119B hybrid Mixture-of-Experts model that combines instruction following, reasoning and agentic coding. It accepts text and images, supports configurable reasoning effort and offers a 256K context window.", "architecture": "Mixture of Experts with 128 experts and 4 active per token", "input": "Text and images → text; function calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mistral-Small-4-OSL" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "publisher model card" ], "listingType": "Open-weight model", "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "note": "" }, { "label": "Mistral AI announcement", "url": "https://mistral.ai/news/mistral-small-4/", "note": "Publisher details and licence." } ] }, { "id": "qwen2-5-coder-32b", "name": "Qwen2.5-Coder-32B-Instruct", "maker": "Alibaba", "categories": [ "code" ], "parameters": "32.5B", "context": "131,072 with YaRN", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Qwen’s 32.5B code instruction model for generation, completion, reasoning and repair. Its model card documents context extension up to 131K tokens using YaRN.", "architecture": "Dense decoder-only Transformer fine-tuned for code", "input": "Text and code → text/code", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen2.5-Coder-32B-Instruct" } ], "benchmarks": null, "benchmarkNotes": "Qwen Team reports Aider 73.7 for code repair, McEval 65.9 for multilingual code generation, and MdEval 75.2 for multilingual code repair. These use task-specific protocols rather than the catalogue’s shared LLM measures.", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct", "note": "" }, { "label": "Qwen Team benchmark blog", "url": "https://qwenlm.github.io/blog/qwen2.5-coder-family/", "note": "Publisher-reported Aider, McEval and MdEval results." } ], "benchmarkHighlights": [ { "label": "Aider · code repair", "value": "73.7" }, { "label": "McEval · multilingual code generation", "value": "65.9" }, { "label": "MdEval · multilingual code repair", "value": "75.2" } ] }, { "id": "qwen2-5-omni-7b", "name": "Qwen2.5-Omni-7B", "maker": "Alibaba", "categories": [ "multimodal" ], "parameters": "7B", "context": "32K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "An end-to-end model that perceives text, images, audio and video and can generate text or speech. Its Thinker-Talker design supports streaming speech output and real-time multimodal interaction.", "architecture": "Thinker-Talker multimodal Transformer with streaming audio output", "input": "Text, images, audio and video → text and speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen2.5-Omni-7B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B", "note": "" } ] }, { "id": "deepseek-r1-distill-qwen-32b", "name": "DeepSeek-R1 Distill Qwen 32B", "maker": "DeepSeek", "categories": [ "text" ], "parameters": "32B", "context": "32K", "license": "MIT", "openness": "Open weights · permissive licence", "description": "A 32B reasoning model distilled from DeepSeek-R1 into the Qwen2.5-32B base. It is intended for mathematics, coding and general problem solving and is released under MIT.", "architecture": "Dense Qwen2.5-32B backbone distilled with reasoning traces", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/DeepSeek-R1" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": 62.1, "lcb": 57.2, "swe": null } }, "benchmarkNotes": "Publisher distilled-model evaluation row: GPQA Diamond and LiveCodeBench pass@1 scores are specific to the 32B distilled checkpoint. The publisher does not list MMLU-Pro or SWE-bench Verified for this checkpoint in that table.", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "note": "" } ], "benchmarkHighlights": [ { "label": "AIME 2024 · pass@1", "value": "72.6%" }, { "label": "AIME 2024 · consensus@64", "value": "83.3%" }, { "label": "MATH-500 · pass@1", "value": "94.3%" }, { "label": "Codeforces rating", "value": "1691" } ] }, { "id": "deepseek-v3", "name": "DeepSeek-V3", "maker": "DeepSeek", "categories": [ "text" ], "parameters": "671B total · 37B active", "context": "128K", "license": "MIT", "openness": "Open weights · permissive licence", "description": "DeepSeek’s 671B MoE language model activates 37B parameters per token. It targets reasoning and coding, uses Multi-head Latent Attention, and supports up to 128K context.", "architecture": "Sparse MoE Transformer with Multi-head Latent Attention and multi-token prediction", "input": "Text → text; tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/DeepSeek-V3" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 75.9, "gpqa": 59.1, "lcb": 37.6, "swe": 42 } }, "benchmarkNotes": "Publisher instruction-model table: MMLU-Pro is 5-shot, GPQA-Diamond is Pass@1, LiveCodeBench is Pass@1, and SWE-bench Verified is reported as resolved. Results use the publisher’s report protocol.", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3", "note": "" } ], "benchmarkHighlights": [ { "label": "HumanEval · Pass@1", "value": "65.2%" }, { "label": "MBPP · Pass@1", "value": "75.4%" }, { "label": "MATH", "value": "61.6%" }, { "label": "AIME 2024 · Pass@1", "value": "39.2%" }, { "label": "IFEval · prompt strict", "value": "86.1%" } ] }, { "id": "mixtral-8x7b", "name": "Mixtral 8x7B Instruct", "maker": "Mistral AI", "categories": [ "text" ], "parameters": "46.7B total · 12.9B active", "context": "32K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Mistral’s instruction-tuned sparse MoE has eight 7B-class experts and routes each token through two. It totals 46.7B parameters with about 12.9B active.", "architecture": "Sparse MoE Transformer with eight experts and two routed per token", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mixtral-8x7B-Chat" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1", "note": "" } ] }, { "id": "mistral-small-3-1", "name": "Mistral Small 3.1", "maker": "Mistral AI", "categories": [ "multimodal" ], "parameters": "24B", "context": "128K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A 24B instruction model for chat, coding and image understanding. It accepts text and images, supports 128K context, function calling and Apache 2.0 use.", "architecture": "Dense Transformer with vision encoder and function-calling support", "input": "Text and images → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Mistral-Small-3.1" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503", "note": "" } ] }, { "id": "phi-4", "name": "Phi-4", "maker": "Microsoft", "categories": [ "text" ], "parameters": "14B", "context": "16K", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Microsoft’s 14B research model focused on reasoning, mathematics and text tasks. It is text-only, supports a 16K context and is released under MIT.", "architecture": "Dense decoder-only Transformer", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/microsoft/phi-4" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Phi-4" } ], "benchmarks": null, "benchmarkNotes": "Microsoft’s model-quality table uses its OpenAI SimpleEval protocol. It reports generic GPQA, not GPQA Diamond; values are shown as reported and are not mapped into the catalogue’s standardized columns.", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/microsoft/phi-4", "note": "" } ], "benchmarkHighlights": [ { "label": "MMLU · SimpleEval", "value": "84.8%" }, { "label": "GPQA · SimpleEval", "value": "56.1%" }, { "label": "MGSM · SimpleEval", "value": "80.6%" }, { "label": "MATH · SimpleEval", "value": "80.4%" }, { "label": "HumanEval · SimpleEval", "value": "82.6%" }, { "label": "SimpleQA · SimpleEval", "value": "3.0%" }, { "label": "DROP · SimpleEval", "value": "75.5%" } ] }, { "id": "phi-4-multimodal", "name": "Phi-4-Multimodal", "maker": "Microsoft", "categories": [ "multimodal" ], "parameters": "5.6B", "context": "128K", "license": "MIT", "openness": "Open weights · permissive licence", "description": "A 5.6B multimodal instruction model built on Phi-4-Mini. It accepts text, images and audio for visual question answering, speech recognition and translation, with 128K context.", "architecture": "Phi-4-Mini language backbone with speech and vision encoders", "input": "Text, images and audio → text; speech tasks", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/microsoft/Phi-4-multimodal-instruct" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Phi-4-Multimodal" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/microsoft/Phi-4-multimodal-instruct", "note": "" } ] }, { "id": "gemma-3-27b", "name": "Gemma 3 27B IT", "maker": "Google DeepMind", "categories": [ "vision", "text", "multimodal" ], "parameters": "27.2B", "context": "128K", "license": "Gemma Terms of Use", "openness": "Open weights · custom or non-commercial licence", "description": "Google DeepMind’s 27B instruction-tuned multimodal model. It accepts text and images, uses a SigLIP vision encoder, supports 128K context and is governed by the Gemma Terms of Use.", "architecture": "Dense decoder Transformer with image encoder and grouped-query attention", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/google/gemma-3-27b-it" }, { "kind": "ollama", "label": "Ollama", "url": "https://ollama.com/library/gemma3" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/google/gemma-3-27b-it" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Gemma-3-27B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 67.5, "gpqa": 42.4, "lcb": 29.7, "swe": null } }, "benchmarkNotes": "MATH 89.0 · HumanEval 87.8 · Bird-SQL 54.4 · LMArena Elo 1338 at launch (publisher card)", "aliases": [ "Gemma 3 27B" ], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/google/gemma-3-27b-it", "note": "" } ], "benchmarkHighlights": [ { "label": "MATH", "value": "89.0%" }, { "label": "HumanEval", "value": "87.8%" }, { "label": "Bird-SQL", "value": "54.4%" }, { "label": "LMArena Elo · launch", "value": "1338" } ] }, { "id": "gemma-3-12b", "name": "Gemma 3 12B IT", "maker": "Google", "categories": [ "vision" ], "parameters": "12.2B", "context": "128K", "license": "Gemma Terms of Use", "openness": "Open weights · custom or non-commercial licence", "description": "A 12B instruction-tuned Gemma 3 model for chat, reasoning and image understanding. It accepts text and images, supports 128K context and uses Google’s Gemma Terms of Use.", "architecture": "Dense decoder Transformer with image encoder and grouped-query attention", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/google/gemma-3-12b-it" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Gemma-3-12B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/google/gemma-3-12b-it", "note": "" } ] }, { "id": "olmo-2-13b", "name": "OLMo 2 13B", "maker": "Ai2", "categories": [ "text" ], "parameters": "13B", "context": "4K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Ai2’s fully open 13B language model, released with training data and code for reproducible research. The base checkpoint has a 4K sequence length and Apache 2.0 licensing.", "architecture": "Dense OLMo 2 decoder-only Transformer", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/allenai/OLMo-2-1124-13B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/allenai/OLMo-2-1124-13B", "note": "" } ] }, { "id": "olmo-2-7b", "name": "OLMo 2 7B", "maker": "Ai2", "categories": [ "text" ], "parameters": "7B", "context": "4K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Ai2’s fully open 7B language model, accompanied by training data and code for reproducible research. The base checkpoint has a 4K sequence length and Apache 2.0 licensing.", "architecture": "Dense OLMo 2 decoder-only Transformer", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/allenai/OLMo-2-1124-7B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/allenai/OLMo-2-1124-7B", "note": "" } ] }, { "id": "falcon-3-10b", "name": "Falcon 3 10B", "maker": "TII", "categories": [ "text" ], "parameters": "10B", "context": "32K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "An instruction-tuned 10B member of TII’s Falcon 3 family, intended for chat, reasoning and code tasks. It supports 32K context under Apache 2.0.", "architecture": "Dense decoder-only Transformer", "input": "Text → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/tiiuae/Falcon3-10B-Instruct" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/tiiuae/Falcon3-10B-Instruct", "note": "" } ] }, { "id": "aya-expanse-32b", "name": "Aya Expanse 32B", "maker": "Cohere For AI", "categories": [ "text" ], "parameters": "32B", "context": "8K", "license": "CC-BY-NC-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "A multilingual generative model trained for instruction following across 23 languages. The 32B checkpoint is for research and multilingual generation; weights are non-commercial.", "architecture": "Dense multilingual decoder-only Transformer", "input": "Text → multilingual text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/CohereLabs/aya-expanse-32b" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/CohereLabs/aya-expanse-32b", "note": "" } ] }, { "id": "command-r7b", "name": "Command R7B 12-2024", "maker": "Cohere For AI", "categories": [ "text" ], "parameters": "8B", "context": "128K", "license": "CC-BY-NC-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "Cohere’s 7B-class Command R model for multilingual chat, retrieval-augmented generation and tool use. It supports 128K context; weights are non-commercial.", "architecture": "Dense decoder Transformer tuned for retrieval and tool use", "input": "Text → text; retrieval and tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/CohereLabs/c4ai-command-r7b-12-2024" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/CohereLabs/c4ai-command-r7b-12-2024", "note": "" } ] }, { "id": "bge-m3", "name": "BGE-M3", "maker": "BAAI", "categories": [ "embed" ], "parameters": "568M", "context": "8K", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Multi-functionality, multi-linguality, multi-granularity embedding model: one checkpoint returns dense, sparse (lexical) and multi-vector (ColBERT) representations for 100+ languages at up to 8192 tokens.", "architecture": "XLM-RoBERTa encoder with dense, sparse and multi-vector retrieval heads", "input": "Text → dense, sparse and multi-vector embeddings", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/BAAI/bge-m3" } ], "benchmarks": { "suite": "embed", "values": { "mteb": null, "ctx": 8192, "dim": 1024, "params": 568 } }, "benchmarkNotes": "Publisher reports MIRACL nDCG@10 around 70 with dense+sparse+multi-vector fusion and MKQA recall; no single MTEB average", "aliases": [], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/BAAI/bge-m3", "note": "" } ], "benchmarkHighlights": [ { "label": "MIRACL nDCG@10 · dense + sparse + multi-vector", "value": "≈70" } ] }, { "id": "e5-mistral", "name": "E5-Mistral-7B-Instruct", "maker": "Microsoft / intfloat", "categories": [ "embed" ], "parameters": "7.1B", "context": "4,096 tokens", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Decoder-based text embeddings fine-tuned from Mistral-7B with synthetic instruction data. Strong English retrieval quality at the cost of a 7B forward pass per document.", "architecture": "Mistral decoder adapted for contrastive text embeddings", "input": "Text with task instruction → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/intfloat/e5-mistral-7b-instruct" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 66.63, "ctx": 4096, "dim": 4096, "params": 7100 } }, "benchmarkNotes": "MTEB (English, 56 datasets) 66.63 (publisher card)", "aliases": [ "E5-Mistral-7B" ], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/intfloat/e5-mistral-7b-instruct", "note": "" } ] }, { "id": "whisper-large-v3", "name": "Whisper-Large-v3", "maker": "OpenAI", "categories": [ "audio" ], "parameters": "1,550M", "context": "Audio processed in 30-second windows", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "State-of-the-art encoder-decoder speech recognition and speech-to-English translation model. Trained on 1M hours of weakly labelled plus 4M hours of pseudo-labelled audio, 128 Mel bins, 99 languages including a new Cantonese token, 10-20% fewer errors than large-v2.", "architecture": "Encoder-decoder Transformer for multilingual speech recognition", "input": "Audio → transcription / translated text", "destinations": [ { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/whisper-large-v3" }, { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/openai/whisper-large-v3" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Whisper-V3-Large-T" } ], "benchmarks": { "suite": "asr", "values": { "wer": 5.78, "rtfx": 470.2, "langs": 99, "params": 1550 } }, "benchmarkNotes": "Open ASR Leaderboard, 11 Sep 2026 English short-form result: published average WER; RTFx measured on H200.", "aliases": [ "Whisper large-v3" ], "provenance": [ "existing catalogue", "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/openai/whisper-large-v3", "note": "" }, { "label": "Open ASR results (11 Sep 2026, English short-form)", "url": "https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/34ee9976c0e916f568301caf4c2341321e2d418b/english_short_latest.csv", "note": "Pinned leaderboard CSV revision; published average WER and H200 RTFx." } ] }, { "id": "mimi", "name": "Mimi Audio Codec", "maker": "Kyutai", "categories": [ "audio" ], "parameters": "Not stated in model card", "context": "12.5 audio tokens/second · 1.1 kbps", "license": "CC-BY-4.0", "openness": "Open weights · permissive licence", "description": "Kyutai’s streaming neural audio codec converts speech to compact audio tokens at 12.5 tokens per second and about 1.1 kbps. The model-card licence is CC-BY-4.0.", "architecture": "Streaming neural audio codec with quantized semantic/acoustic latents", "input": "Speech audio ↔ compressed audio tokens", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/kyutai/mimi" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/kyutai/mimi", "note": "" } ] }, { "id": "stable-diffusion-3-5", "name": "Stable Diffusion 3.5 Large", "maker": "Stability AI", "categories": [ "image" ], "parameters": "8B", "context": "Text-to-image · 1MP-class output", "license": "Stability AI Community License", "openness": "Open weights · custom or non-commercial licence", "description": "Stability AI’s 8B text-to-image checkpoint uses a Multimodal Diffusion Transformer and multiple text encoders. The weights use the Stability AI Community License.", "architecture": "Multimodal Diffusion Transformer with rectified flow and three text encoders", "input": "Text prompt → image", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/stabilityai/stable-diffusion-3.5-large" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/StableDiffusion3.5-L" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/stabilityai/stable-diffusion-3.5-large", "note": "" } ] }, { "id": "flux-1-schnell", "name": "FLUX.1 Schnell", "maker": "Black Forest Labs", "categories": [ "image" ], "parameters": "12B", "context": "Image generation · up to 1024×1024", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A 12B rectified-flow text-to-image model distilled for fast generation, typically in 1–4 steps. It supports prompt-based image creation under Apache 2.0.", "architecture": "Rectified-flow Transformer with parallel self-attention and MLP blocks", "input": "Text prompt → image", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/black-forest-labs/FLUX.1-schnell" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Flux-Schnell-T" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/black-forest-labs/FLUX.1-schnell", "note": "" } ] }, { "id": "smolvlm2", "name": "SmolVLM2", "maker": "Hugging Face", "categories": [ "vision" ], "parameters": "2.2B", "context": "16K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A compact instruction-tuned vision-language model for image understanding, video summaries and visual question answering. The 2.2B checkpoint is designed for efficient deployment under Apache 2.0.", "architecture": "Compact vision-language Transformer with image/video encoder", "input": "Text, images and video → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM2-2.2B-Instruct" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM2-2.2B-Instruct", "note": "" } ] }, { "id": "lfm2-2-6b", "name": "LFM2.5-2.6B", "maker": "Liquid AI", "categories": [ "text" ], "parameters": "2.69B", "context": "131,072", "license": "LFM 1.0", "openness": "Open weights · custom or non-commercial licence", "description": "Liquid AI’s on-device 2.6B agentic model for chat, data extraction, RAG and tool use. The current checkpoint is LFM2.5-2.6B, with 131K context and the LFM 1.0 licence.", "architecture": "Hybrid recurrent/convolutional and attention-based language model", "input": "Text → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/LiquidAI/LFM2.5-2.6B" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/LiquidAI/LFM2.5-2.6B", "note": "" } ] }, { "id": "granite-3-3-8b", "name": "Granite 3.3 8B", "maker": "IBM", "categories": [ "code" ], "parameters": "8B", "context": "128K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "IBM’s 8B instruction model for enterprise chat, retrieval, summarization and code assistance. It supports 128K context and tool use under Apache 2.0.", "architecture": "Dense decoder-only Transformer", "input": "Text → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/ibm-granite/granite-3.3-8b-instruct" } ], "benchmarks": null, "benchmarkNotes": "", "aliases": [], "provenance": [ "existing catalogue" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/ibm-granite/granite-3.3-8b-instruct", "note": "" } ] }, { "id": "qwen3-5-4b", "name": "Qwen3.5-4B", "maker": "Alibaba / Qwen", "categories": [ "text", "vision", "multimodal" ], "parameters": "4B", "context": "262,144", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "A compact 4B multimodal model for text, image and video understanding. It supports 262K native context, can be extended to about 1M tokens and is Apache 2.0 licensed.", "architecture": "Dense multimodal Transformer with hybrid attention", "input": "Text, images and video → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3.5-4B" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen3.5-4B-EL" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 79.1, "gpqa": 76.2, "lcb": 55.8, "swe": null } }, "benchmarkNotes": "Publisher table: MMLU-Pro 79.1 · GPQA Diamond 76.2 · LiveCodeBench v6 55.8; SWE-bench Verified not listed.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B", "note": "" } ] }, { "id": "ministral-3-8b", "name": "Ministral-3-8B-Instruct-2512", "maker": "Mistral AI", "categories": [ "text", "vision", "multimodal" ], "parameters": "8.4B LM + 0.4B vision", "context": "256K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "An 8B-class vision-language instruction model from Mistral 3. It accepts text and images, supports 256K context and tool use, and is Apache 2.0 licensed.", "architecture": "Mistral 3 decoder with vision encoder", "input": "Text and images → text; tool calling", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Ministral-3-8B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": null } }, "benchmarkNotes": "Instruct results: IFEval 87.6 · WildBench 66.8 · Arena Hard 50.9 · MM-MTBench 8.08. The card's GPQA Diamond 66.8 and LiveCodeBench 61.6 are for the separate Reasoning checkpoint.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512", "note": "" } ], "benchmarkHighlights": [ { "label": "IFEval · Instruct checkpoint", "value": "87.6%" }, { "label": "WildBench · Instruct checkpoint", "value": "66.8%" }, { "label": "Arena Hard · Instruct checkpoint", "value": "50.9%" }, { "label": "MM-MTBench · Instruct checkpoint", "value": "8.08" } ] }, { "id": "qwen3-coder-next", "name": "Qwen3-Coder-Next", "maker": "Alibaba / Qwen", "categories": [ "code" ], "parameters": "80B / 3B active", "context": "262,144", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Coding-focused hybrid mixture-of-experts model for tool use and agentic development, released 3 February 2026. Activates 3B of 80B parameters per token, operates in non-thinking mode and supports 370 programming languages. Full weights still need to be resident in memory.", "architecture": "Sparse MoE Transformer with 512 experts and 10 active", "input": "Text and code → text/code; tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/models?q=qwen3-coder-next" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Qwen3-Coder-32B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": 70.6 } }, "benchmarkNotes": "SWE-bench Verified 70.6 (publisher); independent long-context runs report instability beyond ~60K tokens", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next", "note": "" } ] }, { "id": "glm-5-3", "name": "GLM-5.3", "maker": "Z.ai", "categories": [ "text", "code" ], "parameters": "753B", "context": "Up to 1M in publisher evaluations", "license": "GLM-5.3 Model License", "openness": "Open weights · custom or non-commercial licence", "description": "Flagship open-weight text model for coding and agent workflows with configurable reasoning effort. 753B-parameter listing on Hugging Face. Review the model-specific GLM-5.3 licence before deployment; this entry is not MIT or Apache licensed.", "architecture": "Sparse MoE Transformer with IndexShare attention and multi-token prediction", "input": "Text → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/zai-org/GLM-5.3" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/GLM-5.3" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": null } }, "benchmarkNotes": "Weights released August 2026; see publisher card for evaluation tables", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/zai-org/GLM-5.3", "note": "" } ] }, { "id": "glm-5-3-flash", "name": "GLM-5.3-Flash", "maker": "Z.ai", "categories": [ "text", "code", "vision", "multimodal" ], "parameters": "320B total · 18B active", "context": "Up to 1M in publisher evaluations", "license": "MIT", "openness": "Open weights · permissive licence", "description": "First natively multimodal model of the GLM-5 series, released 26 August 2026 on a newly trained base. 320B total / 18B active across 45 layers with 1M context, image and video input, MIT licence. Approaches Claude Opus 4.8 on coding and agentic suites at roughly a tenth of the price.", "architecture": "Sparse MoE with hybrid sparse/linear attention and multimodal input", "input": "Text and images → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "kind": "ollama", "label": "Ollama", "url": "https://ollama.com/library/glm-5.3-flash" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/GLM-5.3-flash" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": null } }, "benchmarkNotes": "Terminal-Bench 2.1 84.3 · DeepSWE v1.1 63.4 · AutomationBench 48.8 · Toolathlon Verified 78.4 · HLE w/ tools 55.3 · NL2Repo 56.3 · AA Intelligence Index 57", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/zai-org/GLM-5.3-Flash", "note": "" } ], "benchmarkHighlights": [ { "label": "Terminal-Bench 2.1", "value": "84.3%" }, { "label": "DeepSWE v1.1", "value": "63.4%" }, { "label": "AutomationBench", "value": "48.8%" }, { "label": "Toolathlon Verified", "value": "78.4%" }, { "label": "Humanity’s Last Exam with tools", "value": "55.3%" }, { "label": "NL2Repo", "value": "56.3%" }, { "label": "AA Intelligence Index", "value": "57" } ] }, { "id": "deepseek-v4-flash", "name": "DeepSeek-V4.1-Flash", "maker": "DeepSeek", "categories": [ "text", "code", "vision", "multimodal" ], "parameters": "552B backbone · 8B active prefill / 16B active decode", "context": "1M", "license": "MIT", "openness": "Open weights · permissive licence", "description": "DeepSeek-V4.1-Flash is the current multimodal Flash checkpoint, released September 2026. It has a 552B backbone, 8B active parameters during prefill and 16B during decode, accepts images and text, and supports 1M context.", "architecture": "40-layer causal encoder-decoder MoE with CSA2, mHC and Engram memory", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "kind": "ollama", "label": "Ollama", "url": "https://ollama.com/library/deepseek-v4-flash" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Deepseek-V4.1-Flash" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 86.2, "gpqa": 88.1, "lcb": 91.6, "swe": 79 } }, "benchmarkNotes": "Official V4.1-Flash release: 10 September 2026. Scores should only be compared when reasoning mode and evaluation protocol match.", "aliases": [ "DeepSeek-V4-Flash" ], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash", "note": "" } ], "old": "DeepSeek-V4-Flash" }, { "id": "deepseek-v4-flash-vision", "name": "DeepSeek-V4-Flash-Vision-Exp", "maker": "DeepSeek", "categories": [ "vision", "multimodal", "experimental" ], "parameters": "284B total · 13B active (V4-Flash base)", "context": "1M family context", "license": "MIT", "openness": "Open weights · permissive licence", "description": "An experimental visual extension of DeepSeek-V4-Flash for image understanding. It adds visual modules to the 284B-total / 13B-active Flash model; DeepSeek labels this checkpoint experimental.", "architecture": "Experimental vision extension of DeepSeek-V4-Flash MoE", "input": "Text and images → text", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Deepseek-Vision-Exp" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": null, "lcb": null, "swe": null } }, "benchmarkNotes": "Third-party (Z.ai comparison table, Aug 2026): Terminal-Bench 2.1 83.9 · DeepSWE v1.1 59.3 · AutomationBench 38.8 · Agents Last Exam 27.3 · MVBench 69.4", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp", "note": "" } ], "benchmarkHighlights": [ { "label": "Terminal-Bench 2.1 · third-party comparison", "value": "83.9%" }, { "label": "DeepSWE v1.1 · third-party comparison", "value": "59.3%" }, { "label": "AutomationBench · third-party comparison", "value": "38.8%" }, { "label": "Agents’ Last Exam · third-party comparison", "value": "27.3%" }, { "label": "MVBench · third-party comparison", "value": "69.4%" } ] }, { "id": "gpt-oss-120b", "name": "GPT-OSS-120B", "maker": "OpenAI", "categories": [ "text", "code" ], "parameters": "117B / 5.1B active", "context": "128K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Open-weight mixture-of-experts reasoning model with 117B total / 5.1B active parameters and native MXFP4 quantisation, so it runs on a single 80 GB GPU. Configurable reasoning effort, full chain-of-thought, tool use and structured outputs under Apache 2.0.", "architecture": "Sparse MoE Transformer with MXFP4 quantized experts", "input": "Text → text; tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/openai/gpt-oss-120b" }, { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/openai/gpt-oss-120b" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/openai/gpt-oss-120b" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/openai/gpt-oss-120b" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/GPT-OSS-120B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": 80.1, "lcb": null, "swe": 62.4 } }, "benchmarkNotes": "MMLU 90.0 · AIME 2025 97.9 (tools) · HLE 19.0 · Codeforces 2622 (tools) · high reasoning effort (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/openai/gpt-oss-120b", "note": "" } ], "benchmarkHighlights": [ { "label": "MMLU · publisher high-reasoning result", "value": "90.0%" }, { "label": "AIME 2025 · tools", "value": "97.9%" }, { "label": "Humanity’s Last Exam", "value": "19.0%" }, { "label": "Codeforces rating · tools", "value": "2622" } ] }, { "id": "gpt-oss-20b", "name": "GPT-OSS-20B", "maker": "OpenAI", "categories": [ "text", "code" ], "parameters": "21B / 3.6B active", "context": "128K", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Smaller sibling of GPT-OSS-120B with 21B total / 3.6B active parameters. MXFP4 weights fit into 16 GB of memory, making it the strongest Apache-licensed reasoning model that runs on a consumer GPU or a laptop.", "architecture": "Sparse MoE Transformer with MXFP4 quantized experts", "input": "Text → text; tool use", "destinations": [ { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/openai/gpt-oss-20b" }, { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/openai/gpt-oss-20b" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/openai/gpt-oss-20b" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/openai/gpt-oss-20b" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/GPT-OSS-120B" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": null, "gpqa": 71.5, "lcb": null, "swe": 60.7 } }, "benchmarkNotes": "MMLU 85.3 · AIME 2025 98.7 (tools) · HLE 17.3 · high reasoning effort (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/openai/gpt-oss-20b", "note": "" } ], "benchmarkHighlights": [ { "label": "MMLU · publisher high-reasoning result", "value": "85.3%" }, { "label": "AIME 2025 · tools", "value": "98.7%" }, { "label": "Humanity’s Last Exam", "value": "17.3%" } ] }, { "id": "kimi-k2", "name": "Kimi-K2-Instruct", "maker": "Moonshot AI", "categories": [ "text", "code" ], "parameters": "1T / 32B active", "context": "128K", "license": "Modified MIT", "openness": "Hosted listing · base model and licence identified", "description": "1-trillion-parameter mixture-of-experts model with 32B active parameters and 384 experts, tuned for agentic tool use and coding in non-thinking mode. Modified MIT licence; native INT4 checkpoints exist. Realistically a multi-GPU or cloud deployment.", "architecture": "Sparse MoE Transformer with 384 experts", "input": "Text → text; tool calling", "destinations": [ { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/moonshotai/kimi-k2-instruct" }, { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/moonshotai/kimi-k2-instruct" }, { "kind": "openrouter", "label": "OpenRouter", "url": "https://openrouter.ai/moonshotai/kimi-k2" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Kimi-K2" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 81.1, "gpqa": 75.1, "lcb": 53.7, "swe": 65.8 } }, "benchmarkNotes": "LiveCodeBench v6 · SWE-bench Verified is agentic single-attempt · MMLU 89.5 · Tau2-bench 66.1 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Hosted assistant listing", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct", "note": "Specifications describe the publisher’s base model; the hosted service can use a different serving configuration." } ] }, { "id": "nemotron-super-49b", "name": "Llama-3.3-Nemotron-Super-49B-v1.5", "maker": "NVIDIA", "categories": [ "text", "code" ], "parameters": "49B", "context": "128K", "license": "NVIDIA Open Model", "openness": "Open weights · custom or non-commercial licence", "description": "Reasoning model NAS-distilled from Llama 3.3 70B down to 49B so that it fits a single H100 or H200 at 128K context. Reasoning can be toggled on or off via the system prompt; released under the NVIDIA Open Model License.", "architecture": "Nemotron-NAS pruned Llama 3.3 Transformer", "input": "Text → text; reasoning and tool use", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/nvidia/llama-3_3-nemotron-super-49b-v1_5" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Nemotron-3-Super" } ], "benchmarks": { "suite": "llm", "values": { "mmlu": 79.53, "gpqa": null, "lcb": null, "swe": null } }, "benchmarkNotes": "Reasoning-on publisher results: MMLU-Pro (CoT) 79.53 · GPQA 71.97 · LiveCodeBench 24.10–25.02 73.58 · MATH500 97.4. GPQA is not labeled Diamond; LiveCodeBench is an older snapshot.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5", "note": "" } ], "benchmarkHighlights": [ { "label": "GPQA · publisher reasoning-on result", "value": "71.97%" }, { "label": "LiveCodeBench · 2024.10–2025.02 snapshot", "value": "73.58%" }, { "label": "MATH-500", "value": "97.4%" } ] }, { "id": "whisper-large-v3-turbo", "name": "Whisper-Large-v3-Turbo", "maker": "OpenAI", "categories": [ "audio" ], "parameters": "809M", "context": "Audio processed in 30-second windows", "license": "MIT", "openness": "Open weights · permissive licence", "description": "A faster distilled Whisper Large-v3 variant for multilingual speech recognition and translation. It uses a four-layer decoder and the 30-second audio-window workflow.", "architecture": "Whisper encoder-decoder with 4 decoder layers", "input": "Audio → transcription / translated text", "destinations": [ { "kind": "groq", "label": "Groq", "url": "https://console.groq.com/docs/model/whisper-large-v3-turbo" }, { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/openai/whisper-large-v3-turbo" }, { "kind": "poe", "label": "Poe", "url": "https://poe.com/Whisper-V3-Large-T" } ], "benchmarks": { "suite": "asr", "values": { "wer": 6.3575, "rtfx": 797.03, "langs": 99, "params": 809 } }, "benchmarkNotes": "Open ASR Leaderboard, 11 Sep 2026 English short-form result: published average WER; RTFx measured on H200.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/openai/whisper-large-v3-turbo", "note": "" }, { "label": "Open ASR results (11 Sep 2026, English short-form)", "url": "https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/34ee9976c0e916f568301caf4c2341321e2d418b/english_short_latest.csv", "note": "Pinned leaderboard CSV revision; published average WER and H200 RTFx." } ] }, { "id": "parakeet-tdt", "name": "Parakeet-TDT-0.6B-v2", "maker": "NVIDIA", "categories": [ "audio" ], "parameters": "600M", "context": "Long-form transcription · up to 24 minutes", "license": "CC-BY-4.0", "openness": "Open weights · permissive licence", "description": "FastConformer-TDT English ASR model with punctuation, capitalisation and word-level timestamps. Transcribes 24-minute segments in one pass and tops the accuracy/speed frontier of the Open ASR Leaderboard at a fraction of Whisper size.", "architecture": "FastConformer encoder with Token-and-Duration Transducer decoder", "input": "English audio → transcription", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2" }, { "kind": "nvidia", "label": "NVIDIA", "url": "https://build.nvidia.com/explore/speech" } ], "benchmarks": { "suite": "asr", "values": { "wer": 4.7025, "rtfx": 6024.67, "langs": 1, "params": 600 } }, "benchmarkNotes": "Open ASR Leaderboard, 11 Sep 2026 English short-form result: published average WER; RTFx measured on H200.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2", "note": "" }, { "label": "Open ASR results (11 Sep 2026, English short-form)", "url": "https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/34ee9976c0e916f568301caf4c2341321e2d418b/english_short_latest.csv", "note": "Pinned leaderboard CSV revision; published average WER and H200 RTFx." } ] }, { "id": "canary-qwen", "name": "Canary-Qwen-2.5B", "maker": "NVIDIA", "categories": [ "audio", "experimental" ], "parameters": "2.5B", "context": "Task-specific audio/text sequence", "license": "CC-BY-4.0", "openness": "Open weights · permissive licence", "description": "Speech-augmented language model: a FastConformer encoder feeds a Qwen3-1.7B LLM decoder, so the same model can transcribe and then summarise or answer questions about the transcript. Debuted at the top of the Open ASR Leaderboard.", "architecture": "FastConformer audio encoder with Qwen decoder", "input": "Audio or text → transcription / translation", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nvidia/canary-qwen-2.5b" } ], "benchmarks": { "suite": "asr", "values": { "wer": 4.4275, "rtfx": 867.09, "langs": 1, "params": 2500 } }, "benchmarkNotes": "Open ASR Leaderboard, 11 Sep 2026 English short-form result: published average WER; RTFx measured on H200.", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nvidia/canary-qwen-2.5b", "note": "" }, { "label": "Open ASR results (11 Sep 2026, English short-form)", "url": "https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/34ee9976c0e916f568301caf4c2341321e2d418b/english_short_latest.csv", "note": "Pinned leaderboard CSV revision; published average WER and H200 RTFx." } ] }, { "id": "seamless-m4t-v2", "name": "SeamlessM4T-v2-Large", "maker": "Meta", "categories": [ "audio" ], "parameters": "2.3B", "context": "Task-specific multilingual speech/text input", "license": "CC-BY-NC-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "Massively multilingual foundation model for speech-to-speech, speech-to-text, text-to-speech and text-to-text translation plus ASR, covering roughly 100 input languages. Non-commercial licence.", "architecture": "UnitY2 encoder-decoder with speech and text modules", "input": "Speech and text → transcription, translation or speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/facebook/seamless-m4t-v2-large" } ], "benchmarks": { "suite": "asr", "values": { "wer": null, "rtfx": null, "langs": 100, "params": 2300 } }, "benchmarkNotes": "Publisher reports FLEURS S2TT BLEU and ASR WER per language; not on the Open ASR Leaderboard", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/facebook/seamless-m4t-v2-large", "note": "" } ] }, { "id": "bark", "name": "Bark", "maker": "Suno", "categories": [ "audio" ], "parameters": "Not stated in publisher card", "context": "Text-to-audio · duration varies by prompt", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Transformer text-to-audio model that produces realistic multilingual speech plus music, background noise and simple sound effects, including nonverbal cues like [laughs] or [sighs]. Three stages: text to semantic tokens, semantic to coarse EnCodec tokens, coarse to fine.", "architecture": "Hierarchical autoregressive semantic and acoustic Transformer stages", "input": "Text prompt → speech, music and sound effects", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/suno/bark" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "No publisher benchmark; vocab 10,000 semantic + 2x1,024 coarse + 6x1,024 fine; detection classifier for Bark audio released", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/suno/bark", "note": "" } ] }, { "id": "musicgen-large", "name": "MusicGen-Large", "maker": "Meta", "categories": [ "audio" ], "parameters": "3.3B", "context": "Text-to-music · up to 30 seconds", "license": "CC-BY-NC-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "Single-stage autoregressive transformer for text-to-music over a 32 kHz EnCodec tokeniser with 4 codebooks at 50 Hz, predicted in parallel via a small inter-codebook delay. Trained on 20K hours of licensed instrumental music; vocals removed with HT-Demucs.", "architecture": "Autoregressive Transformer over EnCodec audio tokens", "input": "Text prompt → music", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/facebook/musicgen-large" } ], "benchmarks": { "suite": "gen", "values": { "fad": 5.48, "kld": 1.37, "clap": 0.28, "sr": 32 } }, "benchmarkNotes": "MusicCaps, released checkpoints: small 4.88 / 1.42 / 0.27 · medium 5.14 / 1.38 / 0.28 · melody 4.93 / 1.41 / 0.27 (chroma cosine 0.44)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/facebook/musicgen-large", "note": "" } ] }, { "id": "audioldm2", "name": "AudioLDM 2", "maker": "CVSSP / Haohe Liu", "categories": [ "audio" ], "parameters": "~1.1B (full)", "context": "Text-to-audio · duration varies by checkpoint", "license": "CC-BY-NC-SA-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "Latent diffusion model for text-to-audio that generates sound effects, music and intelligible speech from one framework using a shared language-of-audio representation. Available in Transformers via diffusers.", "architecture": "Latent diffusion with language-of-audio representation and T5 encoder", "input": "Text or audio prompt → audio", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/cvssp/audioldm2" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 16 } }, "benchmarkNotes": "Paper reports AudioCaps FAD around 1.4 for the full-large checkpoint; not protocol-matched with the MusicCaps figures above", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/cvssp/audioldm2", "note": "" } ], "benchmarkHighlights": [ { "label": "AudioCaps FAD · full-large checkpoint", "value": "≈1.4" } ] }, { "id": "stable-audio-open", "name": "Stable-Audio-Open-1.0", "maker": "Stability AI", "categories": [ "audio" ], "parameters": "1.21B", "context": "Text-to-audio · up to 47 seconds at 44.1kHz stereo", "license": "Stability AI Community License", "openness": "Open weights · custom or non-commercial licence", "description": "Latent diffusion text-to-audio model producing up to 47 seconds of 44.1 kHz stereo sound effects, samples and music loops. Trained only on Creative Commons audio from Freesound and the Free Music Archive; fits in a consumer GPU.", "architecture": "Latent diffusion with T5 text conditioning and autoencoder", "input": "Text prompt → audio", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/stabilityai/stable-audio-open-1.0" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 44.1 } }, "benchmarkNotes": "Publisher reports FD_openl3 and KL_passt on AudioCaps / Song Describer; not comparable to VGGish-FAD", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/stabilityai/stable-audio-open-1.0", "note": "" } ] }, { "id": "xtts-v2", "name": "XTTS-v2", "maker": "Coqui", "categories": [ "audio" ], "parameters": "Not stated in model card", "context": "Text-to-speech · 17 languages", "license": "Coqui Public Model License (CPML)", "openness": "Open weights · custom or non-commercial licence", "description": "Coqui’s multilingual text-to-speech model supports voice cloning from a short reference sample in 17 languages. Its weights are governed by the Coqui Public Model License.", "architecture": "GPT-style acoustic language model with neural vocoder", "input": "Text and reference audio → speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/coqui/XTTS-v2" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "No standard publisher benchmark; community MOS comparisons only", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/coqui/XTTS-v2", "note": "" } ] }, { "id": "kokoro-82m", "name": "Kokoro-82M", "maker": "hexgrad", "categories": [ "audio" ], "parameters": "82M", "context": "Text-to-speech · 24kHz output", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "An 82M open-weight text-to-speech model designed for fast speech generation. The published checkpoint produces 24kHz audio and is Apache 2.0 licensed.", "architecture": "Lightweight style-conditioned neural text-to-speech", "input": "Text → speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/hexgrad/Kokoro-82M" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "Ranked #1 on the TTS Spaces Arena at v0.19 release; no FAD/KLD protocol", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/hexgrad/Kokoro-82M", "note": "" } ] }, { "id": "chatterbox", "name": "Chatterbox", "maker": "Resemble AI", "categories": [ "audio" ], "parameters": "0.5B", "context": "Text-to-speech · duration varies by prompt", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Resemble AI’s open speech model supports zero-shot voice cloning and expressive controls. It accepts text and optional reference audio and returns speech under MIT.", "architecture": "Speech Transformer with speaker conditioning and neural codec decoder", "input": "Text and reference audio → speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/ResembleAI/chatterbox" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "Publisher reports side-by-side preference over ElevenLabs in blind tests; no FAD/KLD protocol", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/ResembleAI/chatterbox", "note": "" } ] }, { "id": "vall-e-x", "name": "VALL-E X", "maker": "Plachta (open reimplementation)", "categories": [ "audio", "experimental" ], "parameters": "Not stated for this checkpoint", "context": "Task-specific · varies by inference implementation", "license": "MIT (community repository)", "openness": "Open weights · permissive licence", "description": "A community implementation of VALL-E X for multilingual zero-shot speech synthesis and cross-lingual voice transfer. This linked checkpoint is community-distributed, not an official Microsoft release.", "architecture": "Autoregressive neural codec language model", "input": "Text and optional speaker audio → speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Plachta/VALL-E-X" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "Community reimplementation; no publisher benchmark", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Community checkpoint and project source", "url": "https://huggingface.co/Plachta/VALL-E-X" } ] }, { "id": "speecht5-tts", "name": "SpeechT5 TTS", "maker": "Microsoft", "categories": [ "audio" ], "parameters": "~144M", "context": "Text-to-speech · 16kHz output", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Microsoft’s unified speech/text Transformer fine-tuned for text-to-speech, with speaker embeddings to control voice characteristics. It produces 16kHz audio under MIT.", "architecture": "Encoder-decoder Transformer with shared speech/text representations", "input": "Text and speaker embedding → speech", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/microsoft/speecht5_tts" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 16 } }, "benchmarkNotes": "Paper reports MOS / CMOS vs. FastSpeech 2; no FAD/KLD protocol", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/microsoft/speecht5_tts", "note": "" } ] }, { "id": "styletts2", "name": "StyleTTS 2", "maker": "Yinghao Aaron Li", "categories": [ "audio" ], "parameters": "Not stated for linked checkpoint", "context": "Task-specific speech synthesis", "license": "MIT (project code)", "openness": "Open weights · permissive licence", "description": "A research text-to-speech system that combines style diffusion with a speech-language-model discriminator. Checkpoint sizes and weight terms vary by release; the linked source is the project repository.", "architecture": "Style diffusion with a speech-language-model discriminator", "input": "Text and reference audio → speech", "destinations": [ { "kind": "project", "label": "GitHub", "url": "https://github.com/yl4579/StyleTTS2" } ], "benchmarks": { "suite": "gen", "values": { "fad": null, "kld": null, "clap": null, "sr": 24 } }, "benchmarkNotes": "Paper: CMOS +0.28 vs. ground truth on LJSpeech; no FAD/KLD protocol", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Project source", "url": "https://github.com/yl4579/StyleTTS2", "note": "The linked HF checkpoint path is not available; this links to the original project source." } ] }, { "id": "qwen3-embedding-8b", "name": "Qwen3-Embedding-8B", "maker": "Alibaba / Qwen", "categories": [ "embed" ], "parameters": "7.6B", "context": "32,768 tokens", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Top-ranked multilingual embedding model built on Qwen3-8B with instruction-aware inputs, 32K context and user-defined output dimensions from 32 up to 4096. 100+ languages including programming languages.", "architecture": "Decoder-based multilingual embedding model with Matryoshka dimensions", "input": "Text with optional task instruction → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3-Embedding-8B" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 70.58, "ctx": 32768, "dim": 4096, "params": 7600 } }, "benchmarkNotes": "MTEB(Multilingual) 70.58 · MTEB(Eng v2) 75.22 · MTEB(Code) 80.68 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3-Embedding-8B", "note": "" } ] }, { "id": "qwen3-embedding-0-6b", "name": "Qwen3-Embedding-0.6B", "maker": "Alibaba / Qwen", "categories": [ "embed" ], "parameters": "0.6B", "context": "32,768 tokens", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Smallest Qwen3 embedding model; same instruction-aware training recipe and 32K context as the 8B version at a size that runs comfortably on CPU or a laptop GPU for local RAG.", "architecture": "Compact decoder-based multilingual embedding model", "input": "Text with optional task instruction → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 64.33, "ctx": 32768, "dim": 1024, "params": 600 } }, "benchmarkNotes": "MTEB(Multilingual) 64.33 · MTEB(Eng v2) 70.70 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "note": "" } ] }, { "id": "nomic-embed-v1-5", "name": "Nomic-Embed-Text-v1.5", "maker": "Nomic AI", "categories": [ "embed" ], "parameters": "137M", "context": "8,192 tokens", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "Fully reproducible long-context embedding model with open weights, open training data and open code. Matryoshka dimensions from 64 to 768 let you trade storage for accuracy on the fly.", "architecture": "Encoder-only Transformer with Matryoshka representation learning", "input": "Text → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 62.28, "ctx": 8192, "dim": 768, "params": 137 } }, "benchmarkNotes": "MTEB (English) 62.28 at 768 dims · LoCo long-context 85.53 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5", "note": "" } ] }, { "id": "jina-v2-base-en", "name": "Jina-Embeddings-v2-base-en", "maker": "Jina AI", "categories": [ "embed" ], "parameters": "137M", "context": "8,192 tokens", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "BERT-sized English embedding model with ALiBi positional encoding that supports 8192-token inputs. A compact choice for long-document retrieval and clustering.", "architecture": "BERT-style encoder with ALiBi long-context attention", "input": "English text → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 60.38, "ctx": 8192, "dim": 768, "params": 137 } }, "benchmarkNotes": "MTEB (English) 60.38 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", "note": "" } ] }, { "id": "gte-qwen2-7b", "name": "GTE-Qwen2-7B-instruct", "maker": "Alibaba-NLP", "categories": [ "embed" ], "parameters": "7.6B", "context": "32,768 tokens", "license": "Apache 2.0", "openness": "Open weights · permissive licence", "description": "General text embeddings on the Qwen2-7B backbone with bidirectional attention and instruction tuning. 32K context and strong multilingual scores across English and Chinese MTEB.", "architecture": "Qwen2 decoder adapted for instruction-conditioned embeddings", "input": "Text with task instruction → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 70.24, "ctx": 32768, "dim": 3584, "params": 7600 } }, "benchmarkNotes": "MTEB (English) 70.24 · C-MTEB 72.05 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct", "note": "" } ] }, { "id": "nv-embed-v2", "name": "NV-Embed-v2", "maker": "NVIDIA", "categories": [ "embed" ], "parameters": "7.85B", "context": "32,768 tokens", "license": "CC-BY-NC-4.0", "openness": "Open weights · custom or non-commercial licence", "description": "Mistral-7B-based generalist embedding model with latent-attention pooling and two-stage contrastive instruction tuning. Held the top MTEB position at release; non-commercial licence.", "architecture": "Mistral-based embedding model with latent-attention pooling", "input": "Text → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/nvidia/NV-Embed-v2" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 72.31, "ctx": 32768, "dim": 4096, "params": 7850 } }, "benchmarkNotes": "MTEB (English, 56 datasets) 72.31 · Retrieval 62.65 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/nvidia/NV-Embed-v2", "note": "" } ] }, { "id": "uae-large-v1", "name": "UAE-Large-V1", "maker": "WhereIsAI", "categories": [ "embed" ], "parameters": "335M", "context": "512 tokens", "license": "MIT", "openness": "Open weights · permissive licence", "description": "Universal AnglE embedding model that optimises an angle objective in complex space to avoid cosine saturation. BERT-large sized, short 512-token window, strong for sentence similarity.", "architecture": "Encoder-only sentence Transformer", "input": "Text → vector embedding", "destinations": [ { "kind": "hf", "label": "Hugging Face", "url": "https://huggingface.co/WhereIsAI/UAE-Large-V1" } ], "benchmarks": { "suite": "embed", "values": { "mteb": 64.64, "ctx": 512, "dim": 1024, "params": 335 } }, "benchmarkNotes": "MTEB (English) 64.64 (publisher card)", "aliases": [], "provenance": [ "supplied catalogue snapshot" ], "listingType": "Open-weight model", "sources": [ { "label": "Publisher model card", "url": "https://huggingface.co/WhereIsAI/UAE-Large-V1", "note": "" } ] } ]