A curated, category-organized registry of open-weight / open-source large language models (LLMs), vision-language models, and related model families.
The goal is to be the definitive place to discover downloadable foundation models you can run locally, fine-tune, or embed in your own products.
Inclusion criteria: Models must ship publicly downloadable weights with a clear open-source or open-weight license. Closed API-only models are out of scope. See CONTRIBUTING.md for the full quality bar.
- General-Purpose LLMs
- Coding LLMs
- Reasoning LLMs
- Small / Fast LLMs
- Multilingual LLMs
- Vision-Language Models
- Fine-Tuned Variants
- Model Families / Hugging Face Collections
- License Notes
- Related Awesome Lists
- Contributing
- License
Large, general-capability text models suitable for chat, RAG, agents, and long-context tasks.
- Llama 3.1 Instruct
Official— Meta's dense instruction-tuned flagship family.- Sizes: 8B, 70B, 405B
- License: Llama 3.1 License
- Llama 3.2 Instruct
Official— Lightweight, multilingual Llama models with vision variants.- Sizes: 1B, 3B
- License: Llama 3.2 License
- Mistral Large 2
Official— Production-grade general-purpose model with a 128K context window.- Sizes: ~123B
- License: Apache 2.0
- Mistral 7B Instruct v0.3
Official— Efficient 7B Apache 2.0 model popular for local deployment.- Sizes: 7B
- License: Apache 2.0
- Qwen2.5 Instruct
Official— Alibaba's dense instruction-tuned family covering a wide parameter range.- Sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B
- License: Apache 2.0 / Qwen License
- Qwen3
Official— Latest Qwen3 flagship MoE with hybrid reasoning modes.- Sizes: 0.6B–235B-A22B
- License: Apache 2.0 / Qwen License
- DeepSeek-V3
Official— MoE frontier model competitive with leading closed-source LLMs.- Sizes: 671B total / 37B active
- License: DeepSeek License
- Gemma 2
Official— Lightweight, high-quality open models by Google DeepMind.- Sizes: 2B, 9B, 27B
- License: Gemma Terms of Use
- Phi-4
Official— Microsoft's state-of-the-art dense model for reasoning and language tasks.- Sizes: 14B
- License: MIT
- Yi-1.5
Official— 01.AI's bilingual base and chat model family.- Sizes: 6B, 9B, 34B
- License: Yi License
Models fine-tuned or pre-trained specifically for code generation, completion, infilling, and software engineering tasks.
- Code Llama
Official— Meta's code-specialized Llama 2 family.- Sizes: 7B, 13B, 34B, 70B
- License: Llama 2 License
- DeepSeek-Coder-V2
Official— MoE code model with strong multi-language performance.- Sizes: 236B total / 21B active
- License: DeepSeek License
- Codestral
Official— Mistral's 22B fill-in-the-middle coding model.- Sizes: 22B
- License: Apache 2.0
- Qwen2.5-Coder
Official— Code-specific Qwen models covering many sizes.- Sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B
- License: Apache 2.0 / Qwen License
- StarCoder2
Official— Transparently trained code generation model by BigCode.- Sizes: 3B, 7B, 15B
- License: BigCode OpenRAIL-M
- Magicoder-S-DS-6.7B
Community— OSS-Instruct distilled coding model.- Sizes: 6.7B
- License: Apache 2.0
- CodeQwen1.5-7B-Chat
Official— Qwen 1.5 code model with long-context support.- Sizes: 7B
- License: Apache 2.0 / Qwen License
- Granite Code
Official— IBM's enterprise-focused code model family.- Sizes: 3B, 8B, 20B, 34B
- License: Apache 2.0
Models optimized for chain-of-thought reasoning, math, science, and complex problem solving.
- DeepSeek-R1
Official— Open reasoning model trained with reinforcement learning, plus distillations.- Sizes: 671B / distilled 1.5B–70B
- License: DeepSeek License
- DeepSeek-R1-Distill-Qwen-32B
Official— Distilled reasoning model combining DeepSeek-R1 outputs with Qwen.- Sizes: 32B
- License: DeepSeek License / Qwen License
- QwQ-32B-Preview
Official— Qwen reasoning model preview with strong math and logic performance.- Sizes: 32B
- License: Apache 2.0 / Qwen License
- Skywork-o1-Open-Llama-8B
Official— Open reasoning model from the Skywork series.- Sizes: 8B
- License: Apache 2.0
- OpenThinker-7B
Community— Reasoning model trained on the Open Thoughts dataset.- Sizes: 7B
- License: Apache 2.0
- Marco-o1
Community— Reasoning model focused on real-world problem solving.- Sizes: 7B
- License: Apache 2.0
Compact models designed for edge, mobile, low-latency, or CPU/GPU-constrained deployments.
- Phi-4-mini-instruct
Official— Small reasoning-capable model from Microsoft.- Sizes: 3.8B
- License: MIT
- Gemma 2 2B IT
Official— Tiny on-device instruction-tuned model.- Sizes: 2B
- License: Gemma Terms of Use
- Qwen2.5-0.5B-Instruct
Official— Sub-billion instruction model for edge use cases.- Sizes: 0.5B
- License: Apache 2.0 / Qwen License
- Llama 3.2 1B Instruct
Official— Edge-friendly multilingual instruction model.- Sizes: 1B
- License: Llama 3.2 License
- SmolLM2 Instruct
Official— Hugging Face small-language-model family.- Sizes: 135M, 360M, 1.7B
- License: Apache 2.0
- TinyLlama-1.1B-Chat
Community— 1.1B chat model trained on 3T tokens.- Sizes: 1.1B
- License: Apache 2.0
- MiniCPM3-4B
Official— 4B model with performance rivaling many 7B models.- Sizes: 4B
- License: Apache 2.0
- Gemma 2B IT
Official— First-generation Gemma on-device instruction model.- Sizes: 2B
- License: Gemma Terms of Use
Models explicitly trained or evaluated for strong performance across many languages.
- Aya 23
Official— Cohere's multilingual family covering 23 languages.- Sizes: 8B, 35B
- License: See model card
- Aya Expanse
Official— Improved multilingual instruction model.- Sizes: 8B, 32B
- License: See model card
- BLOOM
Official— Multilingual LLM developed by the BigScience workshop.- Sizes: 560M–176B
- License: BigScience RAIL License
- XGLM
Official— Meta's multilingual model trained on 30 languages.- Sizes: 7.5B
- License: See model card
- SeaLLM
Official— Southeast Asian language model.- Sizes: 7B
- License: Apache 2.0
- Yi-1.5-34B-Chat
Official— Strong Chinese-English bilingual chat model.- Sizes: 34B
- License: Yi License
Open models that jointly process text and images for captioning, visual QA, OCR, and multimodal agents.
- LLaVA 1.6 34B
Community— Visual instruction-tuned large multimodal model.- Sizes: 34B
- License: Apache 2.0
- Qwen2-VL 7B Instruct
Official— Vision-language model with multilingual OCR and grounding.- Sizes: 7B
- License: Apache 2.0 / Qwen License
- InternVL2 8B
Official— Open-source vision foundation model.- Sizes: 8B
- License: Apache 2.0
- PaliGemma 3B
Official— Google's 3B vision-language model.- Sizes: 3B
- License: Apache 2.0
- MiniCPM-V 2.6
Official— Efficient vision-language model with strong OCR.- Sizes: 8B
- License: Apache 2.0
- Moondream 2
Community— Tiny vision-language model for edge devices.- Sizes: 1.6B
- License: Apache 2.0
- Idefics3 8B
Official— Multimodal model by Hugging Face M4.- Sizes: 8B
- License: Apache 2.0
- Bunny-v1.0-4B
Community— Compact vision-language model from BAAI.- Sizes: 4B
- License: Apache 2.0
Popular community and research fine-tunes built on top of open foundation models.
- Dolphin 2.9 Llama 3.1 70B
Community— General chat fine-tune of Llama 3.1.- Sizes: 70B
- License: Llama 3.1 License
- Nous Hermes 2 Yi 34B
Community— General-purpose Yi fine-tune.- Sizes: 34B
- License: Yi License
- Starling-LM-7B-beta
Community— RLHF-tuned chat model.- Sizes: 7B
- License: Apache 2.0
- OpenChat 3.5
Community— General-purpose conversation fine-tune.- Sizes: 7B
- License: Apache 2.0
- Neural Chat 7B v3
Official— Intel supervised fine-tune of Mistral 7B.- Sizes: 7B
- License: Apache 2.0
- Tülu 3 70B
Official— AI2's instruction-following model suite.- Sizes: 70B
- License: Apache 2.0
- Vicuna 13B v1.5
Community— Early open chat fine-tune of Llama 2.- Sizes: 13B
- License: Llama 2 License
- Zephyr 7B beta
Official— Hugging Face H4 direct preference optimization fine-tune.- Sizes: 7B
- License: Apache 2.0
Curated starting points for browsing full model families and official collections on the Hugging Face Hub.
- Meta Llama 3.1 Collection
Official— All Llama 3.1 base, instruct, and guard models. - Mistral AI Organization
Official— Mistral, Mixtral, Codestral, and embedding models. - Qwen2.5 Collection
Official— Dense and instruction-tuned Qwen2.5 checkpoints. - Google Gemma 2 Collection
Official— Gemma 2 base and instruction-tuned models. - Microsoft Phi-4 Collection
Official— Phi-4 and Phi-4-mini models. - DeepSeek Organization
Official— DeepSeek-V3, DeepSeek-Coder, and DeepSeek-R1 families. - Allen AI OLMo 2 Collection
Official— Truly open-source training-data-and-all language models. - NVIDIA Llama-3.1-Nemotron Collection
Official— NVIDIA's Llama 3.1 Nemotron fine-tunes for helpfulness and reward modeling.
Open-weight LLMs are released under a variety of licenses. Always review the model card before deploying commercially.
- Llama 3.x License — Meta's custom license for Llama 3.1, 3.2, and 3.3 models.
- Apache 2.0 — Permissive license used by Qwen, Mistral Large 2, Gemma 2, and many others.
- MIT — Very permissive license used by Microsoft Phi models.
- BigCode OpenRAIL-M — Responsible AI license used by StarCoder2.
- BigScience RAIL License — License for the BLOOM family.
- DeepSeek License — Custom license for DeepSeek-V3, Coder-V2, and R1.
- Gemma Terms of Use — Google's terms for Gemma models.
- Qwen License — Alibaba's license supplement for some Qwen releases.
- Yi License — 01.AI's custom license for the Yi family.
- Awesome Local LLMs — Tools and resources for running LLMs locally.
- Awesome CLI Coding Agents — Terminal-native AI coding agents and harnesses.
- Cloud GPUs — Comparison site and data for cloud GPU instances.
Read CONTRIBUTING.md for the quality bar, entry format, and PR process.
This list is released into the public domain under CC0-1.0.
Enterprise AI Atlas is maintained by Vibe Coding Agency. We prototype and ship agentic systems, MCP servers, and enterprise AI integrations for teams that need working software fast — without hiring a full AI engineering team.
Free guide: The Non-Technical Founder's Guide to Agentic AI — what agents and MCP servers are, and how to get a system built.