Mistral Large 4 boosts code capabilities but remains behind top Chinese models
Mistral has released the weights of Large 4, its flagship model with about 1 trillion parameters. The tool is strong in code, fast, and cheaper than before, but it is still outpaced by several Chinese models in overall rankings.
Mistral published the weights of Mistral Large 4, its largest model to date, on October 29, 2026. This multimodal model, capable of processing text and images, can be downloaded for deployment on one’s own servers, with a modified open-weight license. On paper, it shows significant progress in programming and defensive cybersecurity. However, in independent rankings, it still lags behind the top Chinese models in overall score.
A giant model, designed to run at home
Mistral Large 4, nicknamed "Le Chonk," has about 1 trillion parameters, of which 52 billion are active per request. In other words, only a subset of the model is used at each generation, which reduces computational load while maintaining very high capacity. The standard version requires eight Nvidia B300 GPUs, according to the inference documentation cited by sources, giving a clear idea of the required muscle.
The company also presents this model as a bet on digital sovereignty. It was trained in Europe, on clean datacenters, and can be run locally rather than in a third-party cloud. For businesses, the advantage is clear: keeping control over data and infrastructure. The downside is more prosaic. It requires space, electricity, and a budget that is not decided over two coffees.
Code improves quickly, overall ranking much less so
Regarding coding, Mistral emphasizes the gains obtained after three additional weeks of reinforcement training. According to the company, Large 4 reaches 48% on the Artificial Analysis Coding Agent Index, tied or nearly tied with GLM-5.3, while consuming 47% fewer tokens. On Terminal-Bench 4.0, the score rises from 28% in the preview to 41% in the final version, placing the model above several open rivals. Mistral also reports achieving 19.7% on FrontierCode 2.0, a private Cognition benchmark focused on end-to-end code quality and the likelihood that a maintainer merges the proposed change.
The picture changes when looking at the overall ranking. On the Artificial Analysis Intelligence Index v5, Large 4 scores 26 points, compared to 32 for DeepSeek V4.1 Flash and GLM-5.3, and 59 for Gemini 4 Carbon. Mistral’s model thus surpasses the median of comparable large open models, but not the Chinese leaders nor the most powerful closed systems from Anthropic and Google. The message is quite simple: good in the workshop, less sharp on the overall track.
Speed and cost gains, with a license that keeps a key aside
The model runs faster than before. Sources indicate a throughput of 161 tokens per second, compared to 111 for the preview. On the API pricing side, Mistral announces $0.70 per million input tokens and $2.10 per million output tokens, with a rate of $0.07 for cached tokens. During the preview, the pricing grid was higher. The company thus shows a significant decrease, even if some Chinese competitors remain cheaper for several uses.
The license follows the same logic: broad openness, but not total. The weights are published under a modified version of the MIT license, free for use, modification, fine-tuning, and commercial deployment. However, actors who want to offer Large 4 as a commercial service in their own cloud must obtain a separate license from Mistral. Openness, yes. Full self-service, no.
Cyberdefense: a good score, but reinforced safeguards
In security, Mistral highlights a score of 51% on the Artificial Analysis Cyber Index, which measures a model’s ability to detect, reproduce, and fix vulnerabilities without breaking the rest of the code. The average rejection rate reaches 96.6% on malicious cyber requests, according to the company. This level of blocking is presented as the highest among tested open models. At the same time, Mistral specifies that some special versions are already tested by partners like Orange Cyberdefense for pentesting, vulnerability hunting, and red teaming exercises, with controlled access and contractual restrictions.
This caution is not decorative. Mistral acknowledges that the model can also facilitate real attacks, hence planned safeguard adjustments in the coming months. Companies interested in a security AI will therefore have to manage a dual movement: more power for defense, but also more control over what is allowed. In this field, the model is not a Swiss army knife; it is more like a toolbox with some pieces kept under lock and key.
Key points
- October 29, 2026: publication of Mistral Large 4 weights.
- About 1 trillion parameters, including 52 billion active.
- 48% on the Coding Agent Index, according to Mistral.
- 26 points on the Artificial Analysis Intelligence Index v5.
- 96.6% rejection rate on malicious cyber requests.
- Orange Cyberdefense tests a version dedicated to pentesting.
By the numbers
- 161 tokens/second — announced throughput for Large 4, according to Artificial Analysis.
- $0.70 — price per million input tokens via the Mistral API.
- $2.10 — price per million output tokens via the Mistral API.
- 19.7% — score on FrontierCode 2.0, according to Mistral.
- 3,800 Nvidia Grace Blackwell GPUs — computing capacity used for training, according to the model sheet.