Open-weight model

The weights are the numbers a model adjusts during training: billions of them, stored in files that can weigh many gigabytes. Families such as Llama, Qwen and DeepSeek publish weights for many of their models. Quantization shrinks those numbers to fewer bits so that a model fits on a consumer graphics card, at some cost in accuracy. Fine-tuning on a few thousand examples, a cheap procedure compared with training from scratch, is how a general model becomes a legal assistant or a coding helper. With the weights and the code that runs them, a person or a company can host the model on its own hardware, fine-tune it for a special task or compress it to run on smaller machines. A closed model offers none of this. It stays on its maker's servers and can only be reached through the maker's own interface, which means the maker sees every request.

Open weights and open source are often confused. Open source in the strict sense asks for more: access to the code, information about the data, and a license that lets anyone use, study, modify and share the work. Many open-weight models come with a license that adds conditions, for example on commercial use or on very large deployments, and the training data is usually kept private. Read the license of the model you plan to build on rather than assume.

The practical effect is competition between hosts. Because the same weights can be served by several companies, a user or an application can choose where a request goes, compare speed and price, and pick a host with a retention policy that suits them. The trade-offs are real too: large models need serious hardware, the weights alone do not reveal what the model was trained on, and anyone can fine-tune them, so two models built on the same weights can behave quite differently: a community tune may be terser, more cautious, specialized in code or tuned for translation. Benchmarks help but mislead: a leaderboard rank rewards particular test formats, so a short trial on your own documents says more than a published score.

Mistral Small 3.2 is an open model from Mistral that NeverNo includes in its Free plan. It is labeled private: requests go to providers chosen because they do not keep them, no account details travel with the message, and your conversations are stored encrypted in your browser, as the private AI page explains. It reads images, takes up to 128,000 tokens of context, writes replies of up to 4,096 tokens and never reasons before it replies, so it answers right away.

On NeverNo, private means exactly this: an open model, served by DeepInfra or by OpenRouter with its zero-retention option, which does not keep requests. The catalog holds several such models, from DeepSeek, GLM, Qwen and Mistral to NeverNo Uncensored. The label rests on those providers' commitments, and NeverNo says plainly on its transparency page that it cannot inspect their servers.