Open-weight model

An open-weight model is one whose weights, the billions of numbers that make up the trained model, are released for anyone to download. Families such as Llama, Qwen, Mistral, Gemma, DeepSeek, and gpt-oss fall into this group. Open weight is narrower than open source: you get the finished model, usually not the training data or the full recipe, and the license decides what you may do with it.

For a builder, the appeal is control. You decide where inference runs, so data can stay on your servers or a customer's. You can pin an exact version that never changes underneath you, fine-tune it for your task, and pay for hardware instead of per token. Many hosting companies also serve popular open-weight models behind OpenAI-compatible endpoints, so you can try one without owning a GPU.

The trade-off is that you inherit the work a provider normally does: serving, scaling, monitoring, and upgrades. The strongest open-weight models have typically trailed the best frontier models on the hardest tasks, though the gap varies by task and you should measure it on yours. Licenses also differ, from Apache 2.0 to custom community licenses with acceptable-use terms you may need to pass on to your users. Read the license and run your own evals before committing.