Mistral has postponed the open release of the weights for its trillion-parameter model.

Mistral Large 4 has one trillion parameters, but right now it cannot be downloaded and run locally. The weights are expected to be released in three weeks, after safety checks. For those who need control over the model and the ability to audit it, this is an important caveat.

For now, only a public endpoint with safety restrictions is available. The company explains that it wants to assess how the open weights could be used for defense rather than attacks. This creates a clear compromise: access control first, then more opportunities for independent study of the model.

Mistral claims it trained the model using its own computing resources: 4,000 NVIDIA GPUs. Target use cases include cybersecurity, finance, and chip design. But these are still stated areas of application, not confirmed results.

The limitations are significant: the open weights have not yet been published, and benchmark results are not ready. So it is too early to judge whether Mistral Large 4 will outperform its competitors or prove strong only at specific tasks. Even after the weights are released, the question will remain of how convincing the safety measures are.

What matters more to you when choosing an AI model: the ability to inspect and run it yourself, or strict access control until the risks have been assessed? Source: AI News & Artificial Intelligence | TechCrunch