Mistral Large 4 weights: what the release and early results show

What changes when model weights are released

Mistral Large 4, nicknamed “Le Chonk,” moved from an API preview to a version whose weights, according to an update dated October 29, had been released. That distinction matters: downloadable weights can make it possible to run a model outside its provider’s API, subject to hardware requirements and the applicable license. The claim about availability and licensing comes from a secondary update. The primary materials available here cover Mistral’s initial announcement and do not, by themselves, confirm that either development took place.

The update describes the weights as being distributed under a modified MIT license. That wording does not necessarily mean the terms are identical to the standard MIT license. The actual legal text determines what uses, modifications and redistribution are allowed. Downloadable files alone also do not establish that a model is open source in every sense, or that commercial use is unrestricted.

Scale and architecture: what one trillion parameters means

Mistral introduced Large 4 as a natively multimodal mixture-of-experts (MoE) model with one trillion total parameters and 52 billion active parameters. An MoE activates a subset of its components to process an input rather than using every parameter at every step. The total and active counts therefore describe different aspects of the model: the first indicates its overall scale, while the second refers to the portion Mistral says is active during an inference step. Neither number alone determines the exact memory required to deploy it.

Mistral’s initial announcement says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company’s own European data centers. It also says the training data covered more than 160 languages. These are manufacturer-reported details about training; they do not show that the model performs equally well in every language or that every later deployment will use the same infrastructure.

Benchmarks: keep different results separate

An update dated October 29 reports a score of 26 for Large 4 in Artificial Analysis’s Intelligence Index version 5. The same update compares that result with 32 for DeepSeek V4.1 Flash and GLM-5.3, and 30 for MiMo-V2.6-Pro. These results are attributed to Artificial Analysis and belong to a later index version. They should not be conflated with the score of 38 that secondary sources associated with the preview under an earlier version. An index revision can change both the tasks and the scoring scale, so scores from different versions are not directly comparable without checking the methodology.

The update also reports that Terminal-Bench 4.0 performance rose from about 28% in the preview to 41% after additional weeks of training. That later figure is presented as a Mistral result, not as an independently verified measurement in the evidence available here. The update attributes a score of 51 on the Cyber Index to Artificial Analysis for the later version. Differences between launch and later results could reflect more training, a different configuration or a changed evaluation. The available material does not establish which factor accounts for each change.

Cybersecurity: useful capability, not a guarantee of protection

Cybersecurity is one of the uses Mistral emphasizes. In its initial announcement, the company reported an 82% score on a test involving reproducing and patching a vulnerability, and a 93% score across Cybench challenges. It also said the model resisted 93.3% of attacks in Lakera’s B3 benchmark. These are results Mistral reported in its announcement. Although the announcement describes some evaluations as independent, the evidence provided here does not include the full protocols, datasets or original test runs needed to reproduce them.

A high score on a vulnerability test does not demonstrate that a system can secure a real network or that every output is safe. Reproducing a flaw and proposing a patch is a bounded task; real security work also depends on the environment, permissions, human review and validation of changes. The distinction between authorized defensive work and malicious requests matters too: enabling vulnerability analysis can support legitimate work while also requiring controls to reduce the risk of misuse.

Cost, speed and agent-based use

A general intelligence evaluation alone cannot determine whether Large 4 is a good fit for a coding agent. In these workflows, a model must interpret a repository, choose tools, run commands and check the result. A plausible-looking response may not correspond to a created file or a completed task. For the preview, Mistral reported scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8%. These are manufacturer-reported figures and depend on the evaluation protocol.

The later update attributes a speed of 161 tokens per second and an average cost of $2.06 per task in the index to Artificial Analysis, compared with $5.54 for the preview. It also says Large 4 generated 110 million output tokens across the task set, down from 150 million in the preview. These figures illustrate why a price per million tokens is not enough to estimate the cost of a real task: output length, the number of steps and the attempts required all matter. They do not guarantee that the model will be cheaper in every application.

Availability and what is not directly confirmed

The October 29 update says the weights were published on Hugging Face and describes a modified MIT license. However, the primary source available here is Mistral’s initial announcement, which only promised to release the weights toward the end of October; it does not include the repository page or license text. The release and license can therefore be described as claims in the update, but they cannot be presented as directly confirmed by the repository materials available here.

Timing also affects how the results should be read. The initial announcement describes a preview; the update refers to a later version and additional training. Scores, pricing and performance can vary with the model version, reasoning mode, provider and execution environment. Before making a deployment decision, identify the exact version, review the applicable license and test the model on your own tasks before relying on it in critical operations.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.