Mistral publishes Large 4’s weights and prices the API at $0.70 per million input tokens
Thursday’s release is the model previewed on 6 October, on Hugging Face as a file of about 1 trillion parameters. Mistral says Terminal-Bench 4.0 is 41 percent, up 13 points, and that a coding-agent score of 48 percent ties GLM-5.3 on fewer tokens. The model card’s aggregate intelligence score is 26. A penetration-testing variant stays behind a vetting gate.
Deep dive5 minIn-depth
Maya Chen5 min read
- Mistral
- Open weights
- Le Chonk
- Cybersecurity

Photo: Nathan Laine / Bloomberg via Getty Images, via FortuneSource page
Mistral on Thursday published the weights of Mistral Large 4, the model it nicknames Le Chonk, and set the API price at $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens. The company’s post says the file is an improved version of this month’s public preview, after more reinforcement learning aimed at coding and at defensive cybersecurity. Reuters, in a story Channel News Asia carries with an October 29 dateline, said the release is meant to help organisations identify and fix security flaws, and that anyone can download the model and run it without an online service. Fortune, at 10:00 a.m. Eastern, dates the preview to 6 October. Mistral’s own documentation page for the model still shows that date and a Public Preview label, beside the new price.
The Hugging Face card, in the repository mistralai/Mistral-Large-4-1T-A52B, describes a mixture-of-experts model of about 1 trillion parameters, with 49 billion active per token and 52 billion once embeddings and output layers are counted. A feature list on the same card rounds the active figure to about 52 billion: 128 routed experts, four of them active per token, and one shared expert. Mistral’s docs page states the size differently, as 52 billion active parameters, 1.05 trillion in total, and a 1.6 billion vision encoder. Both pages give a context window of 1 million tokens. The card says the model reads text and images, returns text, was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European data centres, and is fluent in more than 160 languages, including every official language of the European Union. The API name on the card is mistral-large-4. The post adds three reasoning settings, low, medium, and high, with high as the default. It says low replaces a previous setting called None.
A coding tie, and an aggregate score of 26
On Terminal-Bench 4.0, Mistral says the score rose 13 percentage points from the preview, to 41 percent, which the post calls the lead among open-weight models, with output tokens down 30 percent. Fortune, citing figures Mistral provided, puts the preview score at 28 percent and describes the new result as a tie with China’s GLM-5.3. It also says Anthropic’s Mythos 5.1 scored 60.9 percent on the same test. The move from 28 to 41 is the same arithmetic as a 13-point gain. The two write-ups do not use the same word for where that leaves Large 4 against GLM-5.3. On the Artificial Analysis Coding Agent Index, Mistral’s post prints 48 percent, calls the result effectively tied with GLM-5.3, and says the model uses 47 percent fewer tokens per task on average. Fortune’s characterization of that index is that Large 4 narrowly tops other open models. The figure in the post is 48 percent.
On Cognition’s private FrontierCode 2.0, which Mistral describes as a judgment of whether a repository maintainer would merge the agent’s change, the post gives Large 4 19.7 percent and calls that a lead. In Surge Code’s blind comparisons, Mistral says graders preferred Large 4 to GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, and placed it behind only Claude Opus 5.5 among the models in that test. The same post prints 73.3 percent on Finch, second among open models to Kimi K3; 84.5 percent on BrowseComp, up from 75 percent at a supervised-fine-tuning checkpoint, and higher, it says, than GLM-5.3 and DeepSeek Flash 4.1 1017; and 97.1 percent attack resistance on the B3 agent-security benchmark, against 93.3 percent in the preview. On the Artificial Analysis Cyber Index, for finding a flaw, reproducing it, and patching it without breaking other behavior, Mistral says the score is 51 percent. An average refusal rate on malicious cyber requests is 96.6 percent, which the post calls the highest it measures among open models. Those prompts, it says, are the cyber-related items in JailbreakBench, StrongREJECT, and AgentHarm.
The model card adds a result the coding table does not. On Artificial Analysis Intelligence Index v5, it scores Large 4 at 26 and calls that the highest mark among Western open-weight models. Leading open-weight models from China, the card says, score higher on that aggregate. It says Large 4’s stronger results are in coding and cybersecurity, and that the model surpasses frontier closed models on visual grounding. It does not print the grounding comparison.
The weights are public. One cyber version is not
The licence on the card is the Mistral Ecosystem Open-Weight License, described as a modified MIT licence. The model may be used, modified, fine-tuned, and sold. A business that offers it as a model-as-a-service needs a separate commercial licence from Mistral. The post says the weights are on Hugging Face, for serving with vLLM or SGLang, and that the model is also on the Mistral API, on OpenRouter, on Nvidia-hosted endpoints, and in an Nvidia NIM container. Tina Schuchman, corporate vice president for Microsoft Foundry, is quoted in the post saying Microsoft is the first cloud provider to bring Large 4 to customers, through Foundry. The card’s example vLLM command sets tensor parallelism to eight GPUs. A separate NVFP4 checkpoint, the card says, targets Blackwell GPUs, with a smaller memory footprint and the same performance. The card also says an embedded draft head, which it calls DSpark, measured up to 1.9 times the throughput of the target model alone, without changing the outputs. Mistral’s coding tool Vibe can call mistral-large-4, including against a local server.
The cyber section of the post is narrower than the score table. Mistral says that since the preview it has watched cybersecurity-related use in live traffic, that the model refused black-box attacks against live targets, and that it still sees a potential for uplift in conducting real attacks. It says it is changing its guardrails and will keep updating them. A specialized variant, for penetration testing, vulnerability assessment, and red-teaming, is in preview for design partners including Orange Cyberdefense. Access, the post says, requires vetting and a contract, and testing is limited to systems the customer owns or has explicitly authorized, in controlled environments.
I’m not comfortable living in a world in which open weights do not exist.
— Pierre Stock, Mistral’s vice president of science, to Reuters
Once the model is out, Stock told Reuters, it is impossible to contain access. Reuters’ reading of that line is that a customer cannot be cut off, and that anyone can use the file. Fortune quotes him describing the training hardware as Mistral’s own 10-megawatt cluster in France, “a fraction of what all the competitors use,” and saying the release is “a pivotal moment in the global AI race” and “a turning point for Europe.” Fortune also reports that Mistral has raised roughly $6 billion in three years, including a €3 billion Series D in September led by Samsung that valued the company at more than €21 billion, about $24 billion, and that Anthropic was valued at $965 billion as of August. Those sums are Fortune’s. They are not on the model card. The card’s claim is the file: the weights are public, the aggregate index is not a first-place result, and the penetration-testing variant is still gated.


