The open-source model observatory

The default lightweight embedding baseline for years: tiny, extremely fast, and good enough for most semantic search prototypes.

via Hugging Face

A fully generative text-to-audio model that can produce speech, music, and sound effects, but with less voice-cloning control than dedicated TTS models.

via Hugging Face

A versatile embedding model that unifies dense, sparse, and multi-vector (ColBERT-style) retrieval in one checkpoint across 100+ languages.

via Hugging Face

A retrieval-augmented-generation-focused model with strong citation and tool-use behavior, but licensed for research and non-commercial use only.

via Hugging Face

Cohere's larger RAG/tool-use flagship, strong at grounded enterprise-style workflows but explicitly non-commercial without a separate license.

via Hugging Face

The reasoning-focused sibling of DeepSeek-V3, released under a fully permissive MIT license, that triggered a wave of open reasoning-model distillation.

via Hugging Face

A Qwen2.5-32B base fine-tuned on DeepSeek-R1 reasoning traces, bringing most of R1's reasoning gains to a size that fits on a single GPU.

via Hugging Face

A massive sparse MoE that matched or beat leading closed models on many benchmarks at a fraction of the reported training cost.

via Hugging Face

A large decoder-based embedding model that topped the MTEB leaderboard at release, trading model size for retrieval quality.

via Hugging Face

Widely considered the best open-weight text-to-image model for prompt adherence and detail as of its release, but explicitly non-commercial.

via Hugging Face

A distilled, few-step FLUX.1 variant that is both fast enough for near-real-time generation and fully Apache-licensed for commercial use.

via Hugging Face

The largest Gemma 2 release, competitive with models roughly twice its size on several benchmarks at launch.

via Hugging Face

A very small Google model distilled from larger Gemma teachers, aimed at on-device use where every gigabyte of VRAM matters.

via Hugging Face

A well-rounded mid-size Gemma release that was among the strongest open 9B models at launch, distilled from a larger internal teacher.

via Hugging Face

A fully open vision-language model built on Llama 3 that improved substantially on document/OCR tasks over its Idefics2 predecessor.

via Hugging Face

Meta's first frontier-scale open-weight model, competitive with closed frontier models at release but out of reach for anyone without a GPU cluster.

via Hugging Face

A mid-2024 workhorse dense model that still holds up for general reasoning and coding when 405B-class compute is not available.

via Hugging Face

The default pick for a local 8B chat model: strong instruction following for its size, easy to quantize and run on a single consumer GPU.

via Hugging Face

A small edge/on-device model that trades raw capability for speed and a tiny memory footprint, fine for summarization and simple assistants.

via Hugging Face

A late-2024 refresh that gets close to Llama 3.1 405B quality on many benchmarks at a fraction of the compute cost.

via Hugging Face

One of the earliest and still widely-used open vision-language models, combining a CLIP vision encoder with a Mistral 7B backbone.

via Hugging Face

The classic fully-permissive 7B baseline; not the strongest 7B anymore but has no usage restrictions at all.

via Hugging Face

A joint Mistral/NVIDIA release that packs a 128K context window and a new Tekken tokenizer into a single-consumer-GPU-friendly 12B.

via Hugging Face

A dense 24B tuned to fit on a single high-end GPU, aimed squarely at teams that want Apache-licensed weights without MoE serving complexity.

via Hugging Face

A large permissively-licensed MoE that competes with dense 70B+ models while activating a fraction of its total parameters per token.

via Hugging Face

The model that put sparse mixture-of-experts on the open-weight map: 70B-class quality at 13B-class inference cost, no license strings attached.

via Hugging Face

A strong text/melody-to-music generator from Meta's AudioCraft project, but licensed for research and non-commercial use only.

via Hugging Face

A fully open embedding model, including training data and code, with Matryoshka representation learning for variable-size, resizable embeddings.

via Hugging Face

The larger fully-open OLMo 2 release, giving researchers a reproducible 13B baseline with published data and training code.

via Hugging Face

A fully open model in the truest sense: weights, training data, and training code are all published, not just the checkpoint.

via Hugging Face

A tiny MIT-licensed model built for on-device and edge use, with a 128K context window that is unusually large for its size class.

via Hugging Face

A data-quality-first small model that punches far above its 14B size on math and reasoning benchmarks, fully MIT licensed.

via Hugging Face

A remarkably small diffusion transformer that reaches 4K-capable text-to-image quality at a training cost far below SDXL-class models.

via Hugging Face

The flagship Qwen2-VL tier, competitive with closed multimodal models on visual reasoning, document understanding, and video benchmarks.

via Hugging Face

A vision-language model that handles arbitrary image resolutions and understands video, strong on document/OCR-style tasks for its size.

via Hugging Face

A sweet-spot size for local deployment: noticeably stronger reasoning than 7B while still quantizing down to a single consumer GPU.

via Hugging Face

A strong permissively-licensed mid-size model that often trades blows with much larger dense models on reasoning benchmarks.

via Hugging Face

Alibaba's flagship dense open model at 72B, competitive with Llama 3.1 405B on many benchmarks despite far fewer parameters.

via Hugging Face

One of the strongest fully-permissive 7B models available, with unusually good math and coding for its size.

via Hugging Face

Widely regarded as the strongest fully open code model at its size, often cited as close to GPT-4-class coding performance.

via Hugging Face

A code-specialized 7B that punches well above its weight on code generation and repair benchmarks, and is fully permissive.

via Hugging Face

A reasoning-tuned model that produces long chain-of-thought traces before answering, aimed at math and logic tasks rather than general chat.

via Hugging Face

A carefully data-curated small model that outperforms most other sub-2B models on reasoning and knowledge benchmarks.

via Hugging Face

A sub-billion-parameter model small enough to run in a browser via WebGPU, useful for constrained on-device tasks rather than general chat.

via Hugging Face

Stability's 2024 flagship diffusion transformer, a major quality jump over SDXL with much better text rendering and prompt adherence.

via Hugging Face

A smaller SD3.5 variant tuned to run on more modest consumer hardware while keeping most of the architecture's quality gains.

via Hugging Face

Still the most widely-deployed open text-to-image base model, with by far the largest ecosystem of fine-tunes, LoRAs, and ControlNets.

via Hugging Face

The de facto open standard for speech recognition, with strong multilingual accuracy and no usage restrictions.

via Hugging Face

A pruned-decoder version of Whisper large-v3 that runs several times faster with only a small accuracy tradeoff, good for real-time use.

via Hugging Face

A widely-used voice-cloning TTS model that can clone a voice from just a few seconds of audio across 17 languages, but non-commercial by default.

via Hugging Face