https://mistral.ai/news/mistral-large-4/
Yesterday, Mistral pushed their new frontier model. Large 4 ("Le Chonk") is much better than Large 3 (MoE DeepSeek architecture, but not well received) and their previous flagship model Medium 3.5 (128 B dense).
Large 4 is still in preview mode and the weights are not yet open (expected to be released end of month), but it can be tested through the Mistral API. Currently, there is a 50 % discount; otherwise it would be in the same price class as GLM 5.3. (Large 4 is a MoE with 1 trillion-parameters total and 49 billion active).
So far, it looks very promising in my tests. GLM 5.3 still feels more reliable for agentic tasks, but Mistral definitely caught up to the frontier, making it the best open model outside of China. Mistral also claims stronger cyber capabilities of the model (ahead of GLM 5.3 even). That in itself would be significant (for me, Fable/Opus5 keep refusing all security related questions, so GLM 5.3 is what I am using currently).
Hint: A bit hidden, but with a Mistral Pro subscription (15 Euro per month), there are 255 Euro per month of tokens included. Important: you need a "Vibe API" token, which is not the same as the normal API token. To create one: Code -> Extensions -> look for "Advanced" -> create a token. This can be used in Pi or Mistral Vibe, for instance.
Is Large 4 better than the top closed US models or the top open Chinese models? No, but it is still very encouraging that Mistral is back on the map with a competitive model.