Open source LLM models: open-weight licenses, sizes and hardware

Each entry is a model whose trained weights anyone can download from the developer’s official repository and run on their own hardware. For each one it lists the size, the active parameters for mixture-of-experts models, the context window, the license and whether commercial use is allowed, and how much space the official weights take.

“Open-weight” is not the same as “open-source”. Many of these licenses add conditions, such as a separate agreement above a revenue or user threshold, a required attribution line, or a ban on commercial use. The license columns summarize those conditions; the full license text is linked on each model’s page. For the background, read Open-weight vs. open-source AI.

How to read the sizes. Total parameters decide how much memory a model needs, because all of them have to be loaded, even in a mixture-of-experts model that uses only a fraction for each token. Active parameters decide speed. As a rough guide, a model takes about 2 GB per billion parameters in BF16, 1 GB in 8-bit and a little over 0.5 GB in 4-bit, before the context cache. The memory calculator checks which models fit your GPU and RAM, and Running AI locally explains the trade-offs.

The table covers major releases from Meta, Google DeepMind, OpenAI, Alibaba Qwen, DeepSeek, Mistral AI, Microsoft, NVIDIA, IBM, Moonshot AI, Z.ai, MiniMax, Ai2 and Cohere, and is checked against official sources every week.