AI · 30 October 2026 · 5 min read
Mistral’s weights are out. Its scorecard and Hugging Face’s do not match
On 29 October Mistral posted the weights for Large 4, the trillion-parameter model it previewed on 6 October. The company now puts Terminal-Bench 4.0 at 41 percent. Hugging Face’s card the same day still prints 28.3 percent, and 49 billion active parameters against Mistral’s 52 billion.
Mistral said on 29 October that the weights for Mistral Large 4, which it nicknames Le Chonk, can now be downloaded. The model had been in a public preview since 6 October, when the company promised the weights by the end of the month and described a trillion-parameter system with 52 billion active. The licence is Mistral’s own Ecosystem Open-Weight licence, which the company calls a modified MIT and which limits offering the model as a service. A weight file a lab can host is a different thing from an API key. It is not, by itself, a model anyone has rerun.
The score that moved is Terminal-Bench 4.0, and the two cards do not tell the same story. On 6 October Mistral put the model at 28.3 percent. On 29 October the company said three further weeks of reinforcement learning had lifted that by 13 percentage points, to 41 percent, with 30 percent fewer output tokens, and that this led the open-weight field. Hugging Face’s post the same day still prints 28.3 percent, along with 61.7 percent on DeepSWE and 59.4 percent on SWE-Atlas, which are the preview figures. A reader cannot tell from the Hugging Face page that Mistral believes the terminal score has changed. Use the company’s 29 October post for the 41 percent, and treat it as Mistral’s reading of the benchmark, not as a score this desk has reproduced.
The size does not match either. Mistral’s model card lists about 1.05 trillion parameters in total and 52 billion active, with a context of a million tokens. Hugging Face’s post says about 1 trillion total and 49 billion active, and the same million-token window. Forty-nine and fifty-two are not a rounding difference a caption should hide. The cards do not explain the gap. Prices on Mistral’s own pages are $0.70 per million input tokens, $0.07 for a cached input and $2.10 per million output tokens. A price is what the API charges. It is not what it costs to serve the weights on hardware the buyer already owns.
The rest of the new scorecard should be kept in the same tense. Mistral says that on the Artificial Analysis coding-agent index the model scores 48 percent, about level with GLM-5.3, while using 47 percent fewer tokens per task. Hugging Face still prints a coding-agent index of 49.8 percent and calls the model fifth among open weights on Artificial Analysis. Those are different sentences about the same family of tests. The company also says it leads open models on coding, posts a 51 percent on a cyber index, and a 97.1 percent attack-resistance figure, up from 93.3 percent in the preview. Hugging Face says the weights shipped with support in transformers, vLLM and llama.cpp, in two official variants, with community quantisations alongside. “Best” is Mistral’s adjective. The checkable event is that the weights left the preview. The checkable disagreement is that the host and the maker are not yet showing the same model.