Mistral Releases Le Chonk: 26 Points on the Intelligence Index

Mistral releases the weights of Large 4, its model "Le Chonk," under a modified MIT license. Initial independent results confirm good coding performance and fast execution, but the Artificial Analysis Intelligence Index rates it at 26 points, behind several Chinese open models and far from the leading closed models.

Evaluations: 26 points, rising speed, falling costs

The final high-performance version of Large 4 scores 26 points on the Intelligence Index, compared to 25 for the preview. DeepSeek V4.1 Flash and GLM-5.3 each reach 32 points, while MiMo-V2.6-Pro stands at 30. At the top, Google Gemini 4 Carbon leads with 59 points, just ahead of Claude Opus 5.5, and Kimi K3.1 Pro is the best open model with 46 points, its weights not yet public. In terms of execution, Large 4 increases to 161 tokens per second from 111 previously; only DeepSeek V4.1 Flash and Muse Wonder perform better in the Artificial Analysis comparison. The average cost per task drops to $2.06, down from $5.54 during the preview. It remains lower than GLM-5.3 ($5.01) and Kimi K3.1 Pro ($3.68), but still above DeepSeek V4.1 Flash ($0.60) and MiMo-V2.6-Pro ($0.26). Artificial Analysis still considers its price per token high compared to similar-sized open models. The model generated 110 million output tokens on the index, down from 150 million previously, while the median for comparable models is 100 million. Access to Gemini 4 Carbon remains limited to cybersecurity partners.

Adjusted API pricing and cloud clause license

Via the API, Large 4 is priced at $0.70 per million input tokens and $2.10 for output, after preview rates of $1.36 and $4.18. Mistral had applied a two-week discount to $0.68 and $2.09; the new prices are close to that. For comparison, Kimi K3.1 Pro is set at $2 for input and $10 for output. The model is published under a modified MIT license, allowing free use, modification, and commercial deployment. However, a separate license is required to offer Large 4 as a commercial cloud service. Mistral presents it as a sovereign AI, fully trained in Europe and deployable on private servers. It is also accessible on Microsoft Foundry.

Coding: marked progress on Terminal-Bench and blind tests

After three additional weeks of training, Large 4 improves from 28% to 41% on Terminal-Bench 4.0, ahead of GLM-5.3 at 40%. Mistral reports 48% on the Coding Agents Index of Artificial Analysis, roughly on par with GLM-5.3, with 47% fewer tokens. Claude Code with Claude Opus 5.5 leads this index with 62 points. On FrontierCode 2.0, Large 4 ranks first among open models with 19.7%, while Claude Opus 5.5 scores 46%. According to blind comparisons conducted by Surge, the code from Large 4 is more often preferred over that of GLM-5.3, Kimi K3, and DeepSeek, but it only wins 22% of duels against Opus 5.5. The coding agent rankings confirm the dominance of Claude Code models, with Large 4 and GPT-6.1 Terra positioned in the best value/cost quadrant among 30 models, according to a Pareto efficiency frontier. Comparisons published by Mistral do not account for Kimi K3.1 Pro, whose weights are not public; Moonshot plans to release them by November 4, 2026. Kimi K3.1 Pro is rated at 67% on Terminal-Bench 4.0.

Cybersecurity: 51% on the Cyber Index, but open rivals ahead

Mistral presents Large 4 as an open leader in defense, but its own data shows DeepSeek V4.1 Flash at 57% and MiMo-V2.6-Pro at 56%, ahead of Large 4 at 51% on the Cyber Index of Artificial Analysis, just above GLM-5.3 Flash at 50%. Kimi K3.1 Pro leads with 64%. The additional training only brought one point compared to the 50 points from the preview. In high-performance mode, Large 4 scores 51 for $1.21 per task, while DeepSeek V4.1 Flash 1017 achieves a higher score for a fraction of the cost. Mistral reports an average rejection rate of 96.6% for malicious requests and targets legitimate defensive uses. However, the company acknowledges the risk of abuse and plans to strengthen safeguards in the coming months. Orange Cyberdefense is among the partners testing a special penetration testing-oriented version, with access subject to review and contractual constraints.

Weights published, multimodal capabilities, and hardware requirements

Mistral made the weights of Large 4 publicly available on Hugging Face three weeks after the preview. The model accepts text and image inputs and generates text and code. The SGLang documentation indicates a reference configuration on eight Nvidia B300 GPUs, and a compressed variant on four GB300 GPUs. Initial independent tests highlight good performance in coding.

Training method and announced roadmap

After the preview, Mistral continued reinforcement learning, reporting notable gains in coding and defense. Pierre Stock and Corentin Petit detail an approach based on internal tasks of increasing difficulty, limited by the complexity of tasks the team can design and verify. The company presents the whole effort as a "two-year bet." It also claims progress in office use, ranking just behind Kimi K3 among open models for financial sheet editing and web searching. Mistral states it operates with a fraction of the resources of American and Chinese labs and asserts that with ten times more capital, it would take the lead. Large 4 is intended to serve as a foundation for specialized models, with no comment at this stage on smaller versions. On the Intelligence Index, Large 4 is credited with 26 points and remains behind several open Chinese models and the best closed systems. Mistral emphasizes reduced API prices.