Mistral Large 4 Goes Open: Weights Now on Hugging Face

Mistral AI has released the weights of Mistral Large 4 (ML4) on Hugging Face. The French company unveiled its new flagship model on 6 October but initially kept the weights back. Companies can now download the model and deploy it themselves.

Three more weeks of training and testing

According to the company, the delay was used for three additional weeks of reinforcement learning and for tests with customers and partners. The list of testers is long and mixes industry, finance, defence and tech: ASML, Booz Allen Hamilton, Cisco, Crédit Agricole, Crowdstrike, the European Space Agency, Fortinet, Microsoft, Nvidia and Orange Cyberdefense.

The two-step rollout, first a launch and then the open release, is how Mistral describes the second stage of its offensive. The model that is now available is the version refined during that period.

Why it matters

Open weights matter most to organisations that want control over where and how a model runs. Instead of relying only on a hosted service, they can run ML4 on their own systems, adapt it to their needs and keep sensitive data in-house. That is especially relevant for regulated sectors such as banking and cybersecurity, which are well represented among the testers.

The release also puts Mistral in direct competition with the most capable open models, many of them from Chinese labs. Whether ML4 can match them is the open question. The material available so far does not include independent benchmark results, so a reliable comparison will have to wait for evaluations from third parties and from developers who put the model through real workloads.

What to watch next is how quickly the model is adopted by companies after the download becomes possible, and how it performs in independent comparisons against the leading open models.