Mistral open-sources trillion-parameter Large 4 weights: coding self-test up 13 percentage points, cost per problem down 63%

Beating AI News Flash: French AI company Mistral has officially open-sourced the weights of its flagship model Large 4 (Le Chonk). The model has a total of 1.05 trillion parameters, with approximately 49 billion activated per inference. After opening an API preview on October 6, Mistral conducted three more weeks of reinforcement learning training, and this release is the improved version.

Mistral reported that its Terminal-Bench 4.0 coding score rose from 28% to 41%; it scored 19.7% on FrontierCode 2.0, ranking first among open-source models. On Artificial Analysis's coding Agent index, Large 4 scored 48%, close to GLM-5.3's 49%, while consuming 47% fewer tokens per question on average. However, under Artificial Analysis's own test configuration, Terminal-Bench 4.0 measured only 25%, failing to reproduce the official 41%.

In other tests, the financial table task Finch rose from 67.4% to 73.3%, and the success rate of agents resisting prompt injection attacks increased from 93.3% to 97.1%. The cybersecurity index score was 51%, still below DeepSeek V4.1 Flash's 57%.

The improvement in overall capability was not as large. Artificial Analysis has just upgraded its intelligence index to v5; under the same methodology, Large 4 rose from 25 points in the preview version to 26 points, below GLM-5.3 and DeepSeek V4.1 Flash at 32 points.

But inference efficiency improved markedly: output tokens per question decreased by 39%, and the hallucination rate dropped from 41% to 28%. API pricing was also nearly halved. Based on evaluation pricing and usage as of October 29, the cost per question fell from $5.54 to $2.06, a reduction of 63%.

The model weights have been uploaded to Hugging Face, with FP8 and NVFP4 versions available, supporting self-deployment, fine-tuning, and general commercial use. To operate a commercial model hosting service, separate authorization from Mistral is still required.