Open-Weight Model

An open-weight model is one whose trained parameters, the numeric weights that constitute the model, are published for anyone to download and run. This is the practical foundation of sovereign AI: with the weights in hand, you can run inference entirely on your own hardware, with no API key, no per-token billing, and no prompt leaving your machine. Llama, Mistral, Qwen, and DeepSeek releases are well-known examples.

Open-weight is not the same as open-source

The distinction matters. Open-weight means the parameters are downloadable; it does not guarantee you receive the training data, the training code, or a license permitting any use. Many "open" models ship under restrictive community licenses with usage caps or field-of-use limits. A truly open-source model, by the Open Source Initiative's definition, would also release enough information to recreate it, including data sourcing and training methodology. Most open-weight releases sit in the middle: far more transparent than a closed vendor API, but short of fully reproducible.

Why sovereign operators choose them

Open-weight models let a self-hosting Bitcoiner keep the entire AI stack local. There is no vendor that can deprecate the model, change its behavior, censor outputs, or log queries. You control the sampling parameters, the tokenizer, and the hardware. The trade-off is that you also own the operational burden: VRAM sizing, quantization, and updates are your responsibility.

For those building local AI agents, open-weight models are the natural choice, the same self-custody principle that drives running your own Bitcoin node applies to running your own intelligence.