Open weights

An open-weights model is one where the lab publishes the actual numerical weights — the trillions of floating-point numbers that define what the model does — under a license that lets you download, run, and often fine-tune them. Llama (Meta), DeepSeek, Mistral, Qwen, Gemma, and Phi are the main open-weights families in 2026. You can run them on your own GPU, in a cloud you control, or via inference providers like Together, Fireworks, or Groq. What "open weights" deliberately does not include: the training dataset (often legally radioactive due to copyright concerns), the training code (often proprietary infrastructure), the data filtering and curation pipeline, or the RLHF/post-training data. So you can run the model, but you can't reproduce it. This is why most "open" models in 2026 are technically open-weights, not open-source in the strict sense. Open-weights changed the LLM economics dramatically. Self-hosting Llama 4 or DeepSeek V4 on your own infrastructure can be 5-10x cheaper than calling closed APIs at scale, and the privacy/data-residency story is fundamentally different — your prompts never leave your environment. The catch is operational: you own the GPUs, the deployment, the autoscaling, the safety filtering, and the eval pipeline. For most teams, the API-based closed models are still the default until volume justifies the operational cost. Licenses vary and matter. Llama's license has commercial-use restrictions for very large companies. DeepSeek and Qwen ship under more permissive Apache-2.0-style terms. Gemma has Google-specific use restrictions. Always read the license before building a product on an open-weights model.