Mistral published the 4 big weights, aka Le Chunk. After the preview, the model underwent further training for three more weeks with reinforcement, and the company presented measurements of artificial analysis in the announcement: in Terminal-Bench 4.0 the new version had 41%, the best result among the models with open weights, and the preview, based on Mistral's own measurements, got 28%. In the high-state coding agent index, it nearly catches up with the GLM-5.3, 48% vs. 49%, and produces an average of 47% fewer tokens per task. The dumbbells were released under a different MIT: the model can be used and trained later in commercial products; A separate license is required for companies providing access to a service model. Within MoE, 49 billion are active out of a trillion parameters, the context of a million tokens. Not lazy already GGUF compilation, the most compressed quantum weighs 233 GB. In the API the price is $0.70 per million input tokens and $2.10 per million output tokens, almost the same as the discounted preview I wrote about.

Mistral published the 4 big weights, aka Le Chunk. After the preview, the model underwent further training for three more weeks with reinforcement, and the company presented measurements of artificial analysis in the announcement: in Terminal-Bench 4.0 the new version had 41%, the best result among the models with open weights, and the preview, based on Mistral's own measurements, got 28%. In the high-state coding agent index, it nearly catches up with the GLM-5.3, 48% vs. 49%, and produces an average of 47% fewer tokens per task. The dumbbells were released under a different MIT: the model can be used and trained later in commercial products; A separate license is required for companies providing access to a service model. Within MoE, 49 billion are active out of a trillion parameters, the context of a million tokens. Not lazy already GGUF compilation, the most compressed quantum weighs 233 GB. In the API the price is $0.70 per million input tokens and $2.10 per million output tokens, almost the same as the discounted preview I wrote about.