Mistral previews Large 4, a 1T open-weight model

By VK

Published 7 Oct 2026, 11:27 UTC

Reporting from TechCrunch, WIRED, The Decoder

Mistral Large 4, a 1 trillion parameter model, is in API preview from $0.68 per million tokens in, with weights due by the end of October.

What it means for founders

Read the license before you plan around it. Open weights only help if the terms allow your commercial use. Watch for the license to ship with the weights, expected by October 31, before committing.
Use preview pricing for evaluation, not budgets. At $0.68 in and $2.09 out the API is cheap to test, but the listed rates are double. Model your costs at the full price, then compare against what self-hosting a trillion parameter model would cost in GPUs.
Security tooling is the clearest opening. If your product does vulnerability research, incident response or detection engineering and closed models keep refusing, run your own eval set against ML4 now. Check the refusal rate on your legitimate tasks, not just Mistral's headline scores.
EU data residency is a selling point you can pass on. A Mistral-operated European deployment under European law could shorten procurement with regulated buyers. Ask Mistral for the data processing terms before promising customers anything.

The story

Mistral on October 6 opened a public API preview of Mistral Large 4, a natively multimodal model of roughly 1 trillion parameters that the French lab nicknames "le Chonk," and said the weights will follow by the end of October. Until then the public can reach it only through Mistral Studio, served from Mistral's own European data centers, while security firms, vetted partners and state agencies red-team a version with lighter moderation.

What Mistral Large 4 claims

In its launch post, Mistral says ML4 beats every open-weight model built in the US or Europe and leads in cyber defense, finance and law among open models. The security pitch is pointed: on a benchmark where the model must recreate a genuine flaw in open-source code and then fix it, Mistral reports 82 percent, and says GPT-6 Astra and Claude Opus 5.5 score close to zero because they decline. That result measures refusal policy as much as skill. Mistral also says ML4 turns down malicious cyber requests at a higher rate than rival open models, without explaining how it separates defenders from attackers.

On coding, Mistral ranks ML4 above Qwen3.8 Max and DeepSeek V4 Pro on an Artificial Analysis agentic coding index, useful context if you already use DeepSeek. Independent numbers are more modest. The Decoder reports ML4 scores 38 on the Artificial Analysis Intelligence Index, up from 9 for Mistral Large 3 but well behind the 58 posted by Claude Opus 5.5.

The documentation lists a 1 million token context window. During the preview, a million tokens costs $0.68 in, $2.09 out and $0.07 for cached input, half the listed rates of $1.36, $4.18 and $0.14.

Why the timing matters

The release lands while US access to frontier models is getting less predictable. Cofounder Guillaume Lample told WIRED that owning the model matters even for American companies, since a closed one can be withdrawn at any time. Mistral plans a European deployment it operates end to end under European law and says ML4 can be deployed in a private cloud or on a company's own servers. ML4 also follows Reflection AI's Beam, a 501 billion parameter open-weight model, so the top of the open tier is getting crowded.

What we don't know yet

The license. Mistral says it will publish license terms, architecture details and post-training methods alongside the weights.
How long the discounted preview price lasts.
What changes before release: Mistral says its reinforcement learning run is still going, so final benchmarks may differ from today's numbers.
The hardware needed to self-host it at useful speed.