Mistral opens the weights of its 1-trillion-parameter flagship: coding agents match GLM-5.3 with 47% fewer tokens

Mistral released the weights of its flagship Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal model that had until then been offered only as a public preview. During the preview the model already posted strong results on leading coding benchmarks, but its weights were not open. Mistral says feedback from its developer community, early-access customers and partners was put to work while reinforcement learning continued to scale, driving gains in coding and defensive cybersecurity that are still accelerating with no sign of saturation.

With the weights open, the model's coding performance is placed in direct comparison with other open-weight models. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, effectively tied with the leading open-weight model GLM-5.3. In the same comparison it uses 47% fewer tokens per task on average, and on Terminal-Bench 4.0 it improved by 13 percentage points to 41% over its own public preview while cutting output tokens by 30%.

On Cognition's private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge the agent's change, Mistral Large 4 scores 19.7%, which Mistral describes as leading. That benchmark measures whether a change would be accepted by a maintainer, not merely whether code can be generated. In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred over open-weight models including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated.

The report is limited to Mistral's public account of opening the Mistral Large 4 weights, covering coding and defensive cybersecurity evaluations; the material gives no model card, license terms or specific self-hosting requirements.

Mistral states that coding and security gains are still accelerating with no sign of saturation, which is the releaser's assessment rather than an independent measurement; FrontierCode 2.0 is Cognition's private benchmark, and Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index, though the material does not give the full evaluation detail behind that index.