Mistral AI releases open weights for Mistral Large 4

New weights and coding gains

Mistral AI officially released the trained weights for Mistral Large 4, which the company has nicknamed "Le Chonk." This release follows an initial API preview that occurred earlier in October 2026. By providing the weights, the company allows users to deploy the model within their own private computing environments rather than relying solely on Mistral's hosted services.

The updated model underwent approximately three additional weeks of training, specifically focusing on reinforcement learning. Mistral reports that this extra training period successfully improved the model's coding performance. Specifically, the model achieved a 41% score on Terminal-Bench 4.0, which measures tasks involving terminal operations, marking a notable increase from its previous score of 28.3%.

Beyond coding accuracy, the company noted a 30% reduction in the number of output tokens generated. While a lower token count can lead to faster response times and lower API fees, these savings are not guaranteed to be uniform across all tasks. The actual cost and speed benefits will depend on the specific nature of the workload and the inference settings chosen by the user.

Reasoning and architecture

Mistral Large 4 is built as a Mixture of Experts (MoE) model, which is a design where the system contains many specialized components called experts. Instead of activating the entire model at once, it selects only the necessary experts for each input. The model features a total of 1.05 trillion parameters, but only about 52 billion active parameters are involved in computation during inference.

Users can now adjust the depth of the model's reasoning through three distinct levels: low, medium, and high. The "high" setting is the default for the model, but the "low" setting replaces a previous "None" option. This allows developers to balance the amount of processing power devoted to a task against the complexity of the request to optimize for speed or accuracy.

The model is also multimodal, meaning it can process both text and images as inputs while producing text outputs. Its vision encoder consists of 1.6 billion parameters. While it supports a context length of up to 1 million tokens, Mistral notes that this is a technical specification and does not guarantee that the model will maintain high accuracy at that maximum length.

Comparing global competitors

Despite its gains in coding, Mistral Large 4 does not lead the market in every category when compared to Chinese models. Evaluations from Artificial Analysis show that while Mistral excels in some technical areas, it scores lower on general intelligence benchmarks than models like GLM-5.3. For example, Mistral Large 4 received a score of 26 on the Intelligence Index v5, while both Chinese competitors scored 32.

The cost dynamics also vary depending on which model a company chooses to use. Mistral Large 4 has a lower measured API cost per task than GLM-5.3, but it does not offer a cost advantage over DeepSeek V4.1 Flash 1017. These discrepancies suggest that a model's strength in coding does not always correlate with its overall intelligence score or its price point in a production environment.

Because different benchmarks use different methodologies and reasoning settings, users are encouraged to test models on specific tasks that mirror their own needs. Relying on overall scores can be misleading since the weight of coding versus knowledge work varies between different evaluation frameworks. Companies should evaluate whether the specific coding improvements of Mistral Large 4 justify the costs of its deployment compared to more general-purpose alternatives.

Hardware and licensing

Running a model of this scale in a private environment requires significant infrastructure. Verified configurations for the model include eight NVIDIA B300 GPUs for the FP8 format or four GB300 GPUs for the NVFP4 format. When GPU memory is limited, Mistral recommends that users reduce the context length and the number of concurrent tasks to maintain stability.

The release of the weights provides users with more control over their data and operations, but it also introduces significant overhead. Organizations must manage the underlying hardware, secure the power supply, and assign dedicated staff to maintain the system. This is a major consideration for companies deciding between using a managed API and self-hosting a 1-trillion-parameter model.

The model is distributed under the Mistral Ecosystem Open-Weight License, which is a modified version of the MIT license. This allows for commercial use and modification, but it includes a specific restriction for Model-as-a-Service (MaaS) businesses. Any company that wants to provide third-party access to the model's inference or training must obtain a separate license from Mistral AI.

Safety and cyber capabilities

Mistral AI has also released data regarding the model's performance in cybersecurity contexts. The model received a score of 51% on the Cyber Index, which evaluates its capabilities in the cybersecurity domain. Separately, the company reported an average refusal rate of 96.6% for malicious cyber-related requests across several testing platforms like JailbreakBench and AgentHarm.

The company clarifies that a high refusal rate does not necessarily mean the model is immune to all real-world attacks. Mistral acknowledges that the model could still potentially assist in cyberattacks and states that it will continue to update its safety safeguards. They distinguish these public weights from specialized cyber models offered only to vetted partners.

Furthermore, the company notes that being developed in Europe does not automatically guarantee compliance with all regional regulations. While commercial use is permitted, businesses must still ensure their specific implementations meet their own regulatory requirements. Users are encouraged to evaluate the model's refusal behaviors and safety features as part of their own risk assessment before deployment.