France's Mistral AI released the trained weights for an updated version of its AI model Mistral Large 4 (nicknamed "Le Chonk") on October 29, 2026. The model, which entered API preview on October 6, received roughly three additional weeks of training, and users can now deploy it in their own computing environments. In its announcement, the company said, "We are releasing the weights of the improved Mistral Large 4 today."
Mistral reports that the additional training improved coding performance and reduced output token counts. However, third-party overall evaluations and API cost comparisons show that its advantage over competing Chinese models is not uniform. Companies will need to work out for which tasks Mistral Large 4 is worth adopting, including whether to use an external API or secure their own GPUs, power, and operations staff.
Coding performance after three weeks of additional training
According to Mistral, the updated model, which received additional reinforcement learning, scored 41% on Terminal-Bench 4.0, a benchmark of tasks involving terminal operations. That compares with 28.3% at the October 6 preview announcement, an improvement of about 13 points. The company also reports a 30% reduction in output tokens, but task completion rate and the number of generated tokens are different metrics and should be evaluated separately.
データを表で見る
| Score (%) | |
|---|---|
| Oct. 6 preview | 28.3 |
| Oct. 29 update | 41 |
If fewer output tokens are needed for the same task, API fees and response times could come down. But the 30% figure Mistral reported refers to output token counts, so costs and processing times will not fall by the same percentage for every workload. The actual effect depends on input and output token counts, the nature of the task, and the inference settings used.
On another coding benchmark, the Coding Agent Index, the model scored 48%. Mistral says it achieved results essentially on par with China's GLM-5.3 while cutting average tokens per task by 47%. That 47% is also a token-count comparison and does not translate directly into a cost reduction. In addition, the 49.8% figure from the October 6 announcement used an older index with different evaluation conditions from the current 48%, so the two cannot be used to track performance over time.
Mistral also says it scored 19.7% on FrontierCode 2.0, a non-public coding evaluation from Cognition, the highest among open-weight models in the comparison it presented. However, results published by Artificial Analysis for a test of the same name list different figures. It is unclear whether the two parties evaluated under identical conditions, so numbers from different sources cannot simply be compared to determine rankings.
Reasoning depth can be set to one of three levels: "low," "medium," or "high," with "high" as the default. "Low" replaces the earlier "None" setting, letting users adjust how much processing is devoted to reasoning according to task difficulty and purpose. The announcement does not give a uniform figure for how much token counts or costs fall when the setting is lowered.
How it compares on overall evaluation and API cost
Strong coding scores do not guarantee the same ranking on broad, general evaluations. Artificial Analysis, which evaluates AI models, uses nine tests in its Intelligence Index v5, aggregating four areas (knowledge work, coding, science, and general) at 25% weight each. The two coding tests together account for a quarter of the overall score.
The table below summarizes Intelligence Index v5 scores and weighted-average API task costs as of October 29, 2026, taken from the Artificial Analysis model pages for Mistral Large 4, GLM-5.3, and DeepSeek V4.1 Flash 1017. Reasoning settings follow the figures published on each page.
| Model and reasoning setting | Intelligence Index v5 | Weighted-average API task cost |
|---|---|---|
| Mistral Large 4 high | 26 | $2.06 |
| GLM-5.3 max | 32 | $5.01 |
| DeepSeek V4.1 Flash 1017 max | 32 | $0.60 |
Mistral Large 4 high scores 26 overall, while both Chinese models score 32. Its API cost per task is lower than GLM-5.3's but higher than DeepSeek V4.1 Flash 1017's. Note, however, that "high" and "max" are different reasoning settings for different models, so the results do not necessarily reflect the same compute budget. The task costs shown are weighted averages that reflect not only per-token API prices but also the number of generated tokens, cache usage, and the composition of the evaluation tasks. The numbers may change, and they do not include the cost of running the model in one's own environment.
Against GLM-5.3, Mistral Large 4 scores lower overall but has a lower measured API cost per task. Against DeepSeek V4.1 Flash 1017, the measurements show no advantage for Mistral Large 4 in either overall score or task cost. These differences indicate that stronger coding performance does not necessarily translate into a higher overall evaluation. That said, this comparison does not analyze which non-coding tests contributed most to the score gaps.
Because Intelligence Index v5 uses a different methodology from the earlier v4, scores cannot be compared directly across versions. When selecting a model, it is important to check what is being evaluated and under which reasoning settings, and to test tasks close to your actual use cases rather than relying on overall scores alone.
Per-token API prices should also be considered separately from per-task costs. Mistral Large 4's current pricing is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Actual bills depend on input and output token counts and cache usage. Also, the 161.3 tokens per second that Artificial Analysis reports for Mistral Large 4 high measures output speed once generation has started, not overall processing speed including the wait for the first response.
Running a roughly 1-trillion-parameter model in your own environment
The released Mistral Large 4 is a Mixture of Experts (MoE) model with roughly 1 trillion parameters; Mistral's technical documentation lists the total at 1.05 trillion. An MoE model contains many specialized components (experts) and selects only the ones needed for each input token. Mistral Large 4 chooses 4 of 128 experts and combines them with 1 shared expert.
The number of active parameters involved in computation during inference is 49 billion excluding the embedding and output layers, or 52 billion including them. The difference comes from what is counted, not from different model configurations. However, having about 52 billion active parameters does not mean the model can be run by loading only that portion of the weights into GPU memory.
Documentation for the inference framework SGLang gives verified configurations of eight NVIDIA B300 GPUs for the FP8 format and four GB300 GPUs for the NVFP4 format. These are validated setups, not minimum requirements. The model card describes a separate NVFP4 checkpoint and says inference with vLLM and SGLang is supported. When GPU memory is insufficient, reducing context length and concurrency is recommended.
Mistral Large 4 is also a multimodal model that accepts text and images as input and produces text output. Its vision encoder has 1.6 billion parameters, and it supports context lengths of up to 1 million tokens. However, that 1-million-token figure is a specification limit and does not guarantee consistently high accuracy on inputs of that length. For tasks involving long documents or images, users should verify with their own data that the required accuracy is achieved.
Options for using the model include Mistral's API and Studio, private deployment on Microsoft Foundry, and deployment in your own environment. Releasing the weights gives users a choice in how to manage data and the model, but self-hosting brings burdens such as securing GPUs and power, maintaining the system, and assigning staff.
Cyber capability and refusal of misuse are different measures
Mistral reports a score of 51% on its Cyber Index, which evaluates capabilities in the cybersecurity domain. Separately, it disclosed an average refusal rate of 96.6% for malicious cyber-related requests. That refusal rate averages responses to relevant prompts in JailbreakBench, StrongREJECT, and AgentHarm; it does not mean 96.6% of real-world attacks would be stopped. The ability to handle cybersecurity tasks and the ability to refuse requests that could enable misuse need to be evaluated separately.
Mistral itself acknowledges that the model could assist in real cyberattacks and says it will continue updating its safeguards. Models tailored for specialized cyber uses are offered to vetted partners and should not be equated with the weights released publicly this time. And being developed in Europe does not by itself guarantee safety or compliance with various regulations.
Commercial use is allowed, but model-serving businesses need a separate license
The released weights are covered by the Mistral Ecosystem Open-Weight License, a modified MIT license. It permits using, modifying, further training, and commercially using the model. However, if a user or its affiliates operate a Model-as-a-Service (MaaS) business, they must obtain a separate license before using the model for commercial purposes.
The license defines MaaS as a service that gives third parties access to model inference or further training in a way that lets them substantially control inputs, parameters, and training data. End-user products that embed the model into specific functions or execution environments fall outside this definition, as do services that merely relay access to models hosted by others. In addition, internal use and use through Mistral's official products are not subject to Section 2, which sets out the separate license for MaaS businesses.
Companies considering Mistral Large 4 for their own code development and similar work should first measure completion rates and generated token counts on real tasks, then compare API fees with the compute and operating costs required in their own environment. Providers that offer customers access to the model must also check, beyond performance and cost, whether their service falls under the license's definition of MaaS. The weight release widens the range of operating options, but the value of adopting the model will depend not only on its size or benchmark scores but on the results it delivers in real work and the total cost of operating it.