Skip to content
Komposite News
Markets · Technology · Capital
Thursday, September 10, 2026Global EditionGot a tip?
Open Models

Mistral hands over a trillion parameters

Mistral published the weights of Mistral Large 4 on Thursday, three weeks after a preview open only through its API and to vetted partners. Anyone can now run the model; companies selling it as a hosted service need a separate licence.

Dana Whitfield · House byline
October 29, 20264 min read
A row of black server cabinets with perforated doors, large circular cooling fans and small status displays under yellow overhead cable traysCSIRO, CC BY 3.0, via Wikimedia Commons
Black server cabinets with perforated doors and rows of cooling fans in a GPU computing cluster, in a file photograph.Photograph via Wikimedia Commons

Mistral AI, the Paris model developer, released the weights of Mistral Large 4, a model of about a trillion parameters, on Thursday, so that companies and governments can download and run it themselves, TechCrunch reported. Until now the model, nicknamed "Le Chonk", was available only through Mistral's API and, in a less restricted form, to partners Mistral chose.

The model card on Hugging Face describes a mixture-of-experts model, one that routes each piece of text to a few specialised sub-networks rather than the whole model: 128 routed experts, four active at a time, plus one shared expert. It uses "49 billion active parameters per token (52 billion including embeddings and output layers)", reads text and images, takes up to a million tokens of context and offers low, medium and high reasoning settings. The card's serving examples run across eight GPUs, and a separate compressed version is built for Nvidia's Blackwell chips. Mistral also sells access through its API at $0.70 per million input tokens and $2.10 per million output tokens, and the model is offered through Nvidia, OpenRouter and Microsoft Foundry, Mistral said.

The licence is open with one carve-out. The Mistral Ecosystem Open-Weight License is "a modified version of the MIT license, for both commercial and non-commercial use", the card says, but "operating a Model-as-a-Service business requires a separate commercial license with Mistral." A company can fine-tune the model and build it into its own products; a cloud that wants to rent it out to others has to negotiate. That is narrower than the Apache 2.0 licence Reflection AI promised for its rival Beam model, which Komposite reported on 6 October.

Mistral's case for the model is still cyber defence. It says Large 4 brings "state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally", TechCrunch reported. Pierre Stock, Mistral's vice-president of science, told the publication that access "can't be cut off by definition because the model is open-weighted". Arthur Mensch, the chief executive, said AI should "be a tool for empowerment, not dominance". Mistral says the model refused 96.6 per cent of malicious cyber requests in its testing, which it calls the highest among the open models it measured.

Test (Mistral's figures)Preview, OctoberOpen-weight releaseBest rival Mistral shows
Terminal-Bench 4.0 (agentic terminal tasks)28 per cent41 per centMistral says it leads open-weight models
B3 agent security (attack resistance)93.3 per cent97.1 per centGLM-5.3, 93 per cent
BrowseComp (web research)75 per cent (earlier checkpoint)84.5 per centKimi K3, 91.2 per cent
Artificial Analysis Cyber IndexNot given51 per centDeepSeek-V4.1-Flash, 57 per cent
Surge Code blind preference vs Claude Opus 5.5Not givenWins 22 per cent of comparisonsClaude Opus 5.5 ranked first
Source: Mistral's release post, 29 October 2026. All scores are Mistral's own and were not independently verified. Comparison compiled by Komposite.

Mistral's figures show three weeks of further training. On Terminal-Bench 4.0, which tests agents working at a command line, the released model scores 41 per cent against 28 per cent for the preview, a gain of 13 points; on the B3 agent-security test its attack resistance rose by 3.8 points to 97.1 per cent, by our subtraction. The claims about rivals are where the accounts part. TechCrunch reported Mistral's claim that the model outranks all other open models on the Artificial Analysis Coding Agent Index; Mistral's own post puts it at 48 per cent, "effectively tied" with the Chinese model GLM-5.3, while using 47 per cent fewer tokens per task. On the Artificial Analysis Cyber Index it trails two Chinese models, and in blind human comparisons of code it beat Anthropic's Claude Opus 5.5 in 22 per cent of cases. The benchmarks are Mistral's and were not independently verified.

The release ends an arrangement that drew attention during the preview. Mistral said in early October that until the weights shipped it was red-teaming the model with "cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities", as Komposite reported in Two open models, and no weights yet. Partners quoted in Thursday's announcement include Booz Allen Hamilton, Cloudflare, CrowdStrike and ABN AMRO. Brad Medairy, president of Booz Allen's national cyber business, said Mistral's "support for model customization and data-sovereign deployment also lets us tailor it".

Why it matters: closed developers control who gets their strongest security tools by vetting customers and keeping logs, as Anthropic did when it tiered its cyber verification programme. An open-weight model removes that lever once it is published: whatever safeguards it has are the ones trained into it. Mistral is betting that governments and banks that want to run a capable model on their own hardware will value that control more than they fear what others do with the same file. It joins Meta, which opened the weights of an 816 billion-parameter model this month, in a field Chinese developers have led.

MORE ON THIS STORY