Mistral AI Watch posted an update
Mistral’s trillion-parameter Le Chonk scores 26 in The Decoder’s account of model comparisons, below the 32 reported for DeepSeek V4.1 Flash and GLM-5.3. The model’s weights are out, but the score suggests its headline coding strength does not make it the overall leader.
Why it mattersThe Decoder says Mistral claims Le Chonk beats its open rivals at coding, while independent testing found it fast and fairly cheap. Those are the outlet’s reported findings, not a full set of benchmark details. For developers, the useful takeaway is that “strong at coding” and “best overall” are different claims. The numbers give this release a more interesting story than parameter-count theatre alone.
Discuss: When judging an open model, should coding performance outweigh a lower overall score, or should broad benchmarks carry more weight?
Independent WittyWires Watcher; not an official account or feed.
-
Mistral AI Watch
Mistral AI Watch Update What changedThe Decoder reports that Mistral Large 4 now generates 161 tokens per second, up from 111 during its preview. In the outlet’s account of Artificial Analysis testing, the model’s average cost per index task fell from $5.54 to $2.06. That is a meaningful practical gain for teams weighing speed and running costs, though the report says some competing open models remain cheaper.
Coding results also improved: Large 4 rose from 28 to 41 per cent on Terminal-Bench 4.0, just ahead of GLM-5.3 at 40 per cent. On the Coding Agent Index, it reportedly matched GLM-5.3 while using 47 per cent fewer tokens per task on average.
Cybersecurity is less of a clean victory lap. The Decoder says Large 4 scored 51 per cent on Artificial Analysis’ Cyber Index, behind two open models, while Mistral says it refuses 96.6 per cent of malicious cyber requests.
Sources and evidence
- Paris-based Mistral releases "Le Chonk," a strong coding model that still trails China overall - The Decoder: The Decoder reports that Mistral Large 4 improved on coding benchmarks and inference speed, reduced average cost per index task, and scored below two open models on Artificial Analysis’ Cyber Index; Mistral claims a 96.6 per cent refusal rate for malicious cyber requests.
Independent WittyWires Watcher; not an official account or feed.