Open Weights

Simple Definition

Open weights means an AI company has publicly released the trained parameters — the “weights” — of their model so anyone can download, run, customize, and build on it.

The weights are the result of training an AI on massive amounts of data. They’re what make the model work. Releasing them publicly means you can run the model on your own computer without needing to use the original company’s servers.

Open Weights vs. Open Source

Open source traditionally means you get the full package: the code, the training data, the training process, and the right to use everything freely.

Open weights means you only get the trained model itself (the weights file). The training data and full training code may still be private.

Many models are called “open source” but are actually only “open weights” — like Meta’s Llama models. The distinction matters for researchers who want full reproducibility, but for most practical uses, open weights is what you need.

Why Open Weights Matter

You can run it locally — download the model, run it on your own hardware. Local LLMs depend on open weights models.

You can fine-tune it — train the model further on your own data to specialize it for your use case.

No usage restrictions or costs — within the license terms, you can use it how you want.

Transparency and research — researchers can study the model and understand its behaviors.

Popular Open Weights Models

Llama 3 (Meta) — widely used, available in several sizes
Mistral / Mixtral (Mistral AI) — efficient models popular for local use
Phi-3 / Phi-4 (Microsoft) — small but capable models
Gemma (Google) — designed for efficiency and local deployment
DeepSeek-R1 — reasoning-capable open weights model

License Considerations

Open weights models come with different licenses. Some allow commercial use; others are restricted to non-commercial research. Always check the license before building products on top of an open weights model.