55 open-source models indexed
Open source AI models: the directory.
Open source AI models compared by license, parameters, benchmarks, VRAM and inference pricing for Llama, Mistral, Qwen, DeepSeek, Gemma, Phi and more. Pick the right open-weights model for your use case.
“Open source” is used broadly here: models under Apache 2.0 or MIT are listed alongside source-available models such as Llama and Gemma, and every model is tagged with its license. See the license comparison and the FAQ for what that means for commercial use.
Best models for…
All use cases →Curated picks for the six workloads we get asked about most.
The best open-source AI models to run locally on consumer hardware in 2026. Picks span Apple Silicon and RTX 4090 setups.
Open-source models that excel at code generation, completion, and review. Picks span on-device 7B models through frontier-class 30B+ specialists.
Open-source models that run well on consumer hardware (RTX 4090, Apple Silicon, even laptop iGPUs). Picks balance quality with VRAM footprint.
Models tuned for retrieval-augmented generation: long context, strong instruction following, native citation behaviour where possible.
Open-weights multimodal models that can read images, do OCR-adjacent tasks, and reason over diagrams and screenshots.
Models that generate long internal chains-of-thought before answering — designed for math, code, and multi-step problem solving.
Sub-4B models that run on phones, laptops, and embedded devices. Optimised for memory footprint and tokens-per-second on integrated GPUs.
Recently released
Full timeline →Moonshot AI's 1-trillion-parameter mixture-of-experts (32B active per token). Trained on 15.5T tokens with a heavy emphasis on tool-use and agentic behaviour. Modified-MIT licence with an attribution clause for very-large deployments. Exceptional at long-horizon agent tasks; benchmarked well against Claude Sonnet on SWE-bench Verified.
- Context
- 128K
- License
- kimi
- VRAM Q4
- 600 GB
The flagship Qwen 3 release: a 235B-total MoE with 22B active parameters per token. Competitive with DeepSeek V3 and Llama 4 Maverick on reasoning benchmarks while being smaller total. Apache 2.0 — one of the most permissively licenced frontier-class models.
- Context
- 128K
- License
- apache-2-0
- VRAM Q4
- 141 GB
32B sweet-spot Qwen 3, Apache 2.0. Reasoning-mode toggle inherited from smaller siblings; strong on math, code and agentic tool use. Fits on a single H100 in fp16 and on a 4090 at Q4.
- Context
- 33K
- License
- apache-2-0
- VRAM Q4
- 19.2 GB
The April 2025 refresh of Qwen at 8B. Native mixed-mode reasoning: the model can 'think' before answering when triggered, or answer directly for simple queries — configurable per request. Apache 2.0. A strong upgrade over Qwen 2.5 7B on math and code, with much better instruction following.
- Context
- 33K
- License
- apache-2-0
- VRAM Q4
- 4.8 GB
Cheapest hosted right now
Cost calculator →Featured models
See all →Moonshot AI's 1-trillion-parameter mixture-of-experts (32B active per token). Trained on 15.5T tokens with a heavy emphasis on tool-use and agentic behaviour. Modified-MIT licence with an attribution clause for very-large deployments. Exceptional at long-horizon agent tasks; benchmarked well against Claude Sonnet on SWE-bench Verified.
- Context
- 128K
- License
- kimi
- VRAM Q4
- 600 GB
671B-parameter MoE model with 37B active per token. Trained for roughly $5.6M of compute — a landmark in cost-efficient frontier training. Frontier-class quality at a fraction of the cost of the closed proprietary frontier. The DeepSeek licence permits commercial use with limited restrictions on military and unlawful applications. Running V3 yourself requires serious hardware (8× H100 at fp8); most teams will use it via the DeepSeek API or providers like Together.
- Context
- 128K
- License
- deepseek
- VRAM Q4
- 402.6 GB
Reasoning model trained with reinforcement learning on top of DeepSeek V3-Base. MIT licence — even the weights are unrestricted, making R1 the most permissively-licensed frontier reasoning model. Generates long internal chains-of-thought before answering, trading latency for accuracy on math, code, and reasoning benchmarks. Distilled variants (e.g. R1 Distill Llama 70B) recover most of the quality at much smaller scales.
- Context
- 128K
- License
- mit
- VRAM Q4
- 402.6 GB
Meta's July 2024 flagship — the first open-weights model at 405B parameters. Trained on 15T tokens with 128K context. Rivals GPT-4o on many academic benchmarks and set the ceiling for open-weights quality for most of 2024. Running it self-hosted requires serious hardware (8× H100 at fp8 or multi-node at fp16); most users will run it via a hosted provider (Together, Groq, Fireworks). Llama 3.3 70B closed most of the practical gap at a fraction of the cost, so 405B is now most useful when 70B specifically hits its ceiling.
- Context
- 128K
- License
- llama-3
- VRAM Q4
- 243 GB
Hybrid Mamba-Transformer-MoE model with native 256K context (effective beyond 140K). 94B active parameters out of 398B total. The state-space-model layers give it linear-time scaling with sequence length, making it interesting for very long contexts. Licensed under AI21's open model licence, which permits most commercial use.
- Context
- 256K
- License
- jamba-open
- VRAM Q4
- 238.8 GB
NVIDIA's reward-modelling research vehicle. Trained primarily to be a synthetic-data-generation specialist rather than a chat-first model. Useful for teams building instruction-tuning datasets at scale.
- Context
- 4K
- License
- llama-3
- VRAM Q4
- 204 GB
Browse by family
Cohere's enterprise-oriented model family. Command R+ targets retrieval-augmented generation workflows.
Databricks' 132B mixture-of-experts (12 of 16 experts active per token) built on the MegaBlocks stack. Released under the Databricks Open Model Licence.
Hangzhou-based lab known for highly efficient MoE training. DeepSeek V3 and R1 set new bars for open reasoning and coding.
Technology Innovation Institute (UAE). Falcon Mamba pioneered state-space-model open releases.
Google's open-weights model family derived from the same research as Gemini. Strong performance at small scales.
xAI's open-weights releases. Grok 1 and Grok 2 weights have been published under Apache 2.0.
Nous Research's community-driven fine-tune series. Hermes 3 sits on top of Llama 3.1 base weights, tuned for very strong tool use and steerable roleplay behaviour.
Hybrid Mamba/Transformer architecture with very long native context windows.
Beijing-based lab founded by ex-Tsinghua researchers. Kimi K1.5 and K2 push very long context and reasoning; K2 is a trillion-parameter MoE with strong tool-use behaviour.
Meta's open-weights LLM family. Llama 3 introduced 128K context and strong instruction tuning across 1B–405B parameter scales.
Paris-based lab whose dense and mixture-of-experts open-weights models popularised sliding-window attention and high efficiency.
NVIDIA's research and instruction-tuning effort built on top of Llama base models, with strong RLHF and reward-modelling work.
AI2's fully-open initiative: weights, data and training code all under permissive licences.
Microsoft's small-language-model line, focused on data-quality-driven training. Phi-4 punches well above its weight class.
Alibaba's Qwen series. Qwen2.5 delivers competitive frontier-class performance across sizes from 0.5B to 72B with permissive licensing.
Stability AI's general-purpose text models.
01.AI's bilingual (Chinese/English) open-weights series including vision variants.
MMLU leaderboard — top 5
Full leaderboards →- 90.80
- 89.50
- 88.50
- 87.30
- 87.10
Why open-source models matter
Open-weights models let you run inference locally, fine-tune on private data, and avoid vendor lock-in. Open Source AI Models tracks every model with public weights so you can pick on facts, not marketing.