Open weights

With the weights in hand, you can run the model on your own computer or server, adapt it to a specific use, and avoid depending on a single provider.

Open weights is not open source

Under the Open Source Initiative definition, open source AI must release the weights, all training and inference code, and enough information about the data to rebuild an equivalent model. Most models called “open” release only the weights. That is why the more precise term is open weights.

The license decides what you can do

• Permissive (Apache 2.0, MIT): free commercial use. As of November 2026: DeepSeek V4 (MIT), Gemma 4, Mistral 3, gpt-oss, the smaller Qwen 3.8 sizes and Tencent Hy3 (Apache 2.0).
• Custom, with restrictions: Llama 4 (companies above 700 million users need a separate license), Qwen3.8-Max, Kimi K3.1 Pro and GLM-5.3 (conditions based on revenue or company size).

Before using one at work, read the license of the exact model, not the family.

Why consider it

Privacy (data stays in your infrastructure), predictable cost and freedom to adapt. In exchange, you handle inference yourself, and the best open models still trail the closed frontier ones.