Match model size to hardware
A rough rule: a model needs about 2 bytes per parameter in 4-bit quantization. An 8B model needs ~16GB VRAM; a 70B model needs ~140GB. Start small and scale up only when needed.
Open-Source AI Model Matrix
We earn commissions when you shop through the links below.
Searchable reference of open-weights LLMs: Llama, Qwen, Mistral, DeepSeek, Gemma, Phi. Compare params, context window, license, and use case. 100% client-side.
All data is local.
Model specifications are embedded in the page. Nothing is uploaded or fetched from the network.
16 models
| Model | Params | Context | License | Best For |
|---|---|---|---|---|
| Llama 3.1 8B | 8B | 128k | Llama Community | General |
| Llama 3.1 70B | 70B | 128k | Llama Community | General |
| Llama 3.1 405B | 405B | 128k | Llama Community | Reasoning |
| Qwen 2.5 7B | 7.2B | 128k | Apache 2.0 | Coding |
| Qwen 2.5 72B | 72B | 128k | Apache 2.0 | General |
| Mistral 7B | 7.3B | 32k | Apache 2.0 | General |
| Mistral Large 2 | 123B | 128k | Mistral Research | Reasoning |
| DeepSeek-V3 | 671B | 128k | MIT | Coding |
| DeepSeek-R1 | 671B | 128k | MIT | Reasoning |
| Gemma 2 2B | 2.6B | 8k | Gemma Terms | General |
| Gemma 2 9B | 9.2B | 8k | Gemma Terms | General |
| Gemma 2 27B | 27B | 8k | Gemma Terms | Coding |
| Phi-3-mini | 3.8B | 128k | MIT | General |
| Phi-3-medium | 14B | 128k | MIT | Reasoning |
| Command R7B | 7B | 128k | CC-BY-NC | General |
| Command R+ | 104B | 128k | CC-BY-NC | Reasoning |
Table is sortable and filterable. Use the controls above to narrow results by license, use case, or model family.
Frequently Asked Questions
- An open-weights model is an AI model whose trained parameters are publicly available for download and local use, though the license may restrict commercial use (e.g., Llama Community, CC-BY-NC).
- Not all of them. Licenses vary: Apache 2.0 and MIT are permissive for commercial use; Llama Community and CC-BY-NC have restrictions. Check each model's license before deploying commercially.
- Yes—smaller models (7B–14B) run well on consumer GPUs or even CPUs with quantization. Larger models (70B+) typically require multiple GPUs or inference servers.
- For coding: Qwen 2.5 or DeepSeek-V3. For general chat: Llama 3.1 8B or Gemma 2 9B. For reasoning: Llama 3.1 405B or DeepSeek-R1. Consider your hardware's VRAM when choosing size.
- Yes. Once the page is loaded, the entire model matrix runs locally. No internet connection is required.
Choosing the Right Open-Source LLM
Context window matters
If you're feeding long documents or codebases, prioritize models with 128k context (Llama 3.1, Qwen 2.5, DeepSeek). For short tasks, 8k–32k is sufficient and cheaper to run.
License before download
Commercial projects should avoid restrictive licenses (CC-BY-NC, Llama Community for >700M users). MIT and Apache 2.0 are safest for business use.
Quantization trade-offs
Quantizing a model (4-bit, 8-bit) reduces VRAM and increases speed with minimal quality loss. Start with 4-bit GGUF or EXL2 formats for local inference.