Mistral Previews 1T-Param Open-Weight Large 4 'Le Chonk'
TL;DR
- Mistral's mixture-of-experts design activates only 49B of 1T parameters per inference token, keeping serving costs closer to a mid-size dense model despite the headline count.
- Mistral's own blog cites 3,800 Grace Blackwell GPUs while CNBC and the primary alert report 4,000, a minor rounding difference but worth watching in subsequent corrections.
- VentureBeat's independent benchmark cross-check puts ML4 at 62% on DeepSWE against GLM-5.3's reported 69%, confirming Chinese open-weight models hold a coding benchmark edge at launch.
Mistral said on Monday it had trained a 1 trillion-parameter mixture-of-experts model, 49 billion active, on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters. The French lab is calling it Mistral Large 4, or in its own words, "Unofficially ML4, very officially: le Chonk."
A public preview went live on Mistral's API today, priced at $1.36 per million input tokens and $4.18 per million output. "We will release the weights by the end of the month," the company says. ML4 supports 160+ languages, including every official EU language.
The competitive frame is unusual. Mistral benchmarks ML4 head-to-head against DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3 and GLM-5.3, and claims the model is "significantly outperforming any open-weight model developed in the US or Europe." The write-up pitches cybersecurity as a wedge, reporting 93% on Cybench and 82% on a vulnerability reproduction and patching test. The company notes that "several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse to perform the task."
"ML4 is the first milestone on the roadmap funded by our €3 billion Series D," Mistral writes, referring to the raise it calls the largest equity round ever by a European technology company. That round closed in September with Samsung leading. The post is signed "By Mistral"; no executive is quoted by name.
What others are reporting
-
CNBC Read →
Tier-1 business press coverage anchors on sovereign compute: training ran entirely on Mistral's own European data centers, framing the release as an independence play from US and Chinese cloud infrastructure.
ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over two months that were deployed in Mistral's own data centers in Europe.
-
VentureBeat Read →
Cross-references Mistral's benchmark claims against public leaderboards, finds the Chinese coding lead persists, and flags Mistral's pivot toward infrastructure and customization services as the real business story.
ML4 is at the frontier of open weight models. -- Guillaume Lample, Mistral co-founder and chief scientist
-
Simon Willison's Weblog Read →
Independent practitioner assessment: Artificial Analysis score of 38 versus Large 3's 9 shows a dramatic internal leap, but hands-on testing and leaderboard data place ML4 roughly six months behind the proprietary frontier.
Shared on Bluesky by 2 AI experts
Originally reported by mistral.ai
Read the original article →Original headline: Mistral Releases 1T-Param Open-Weight Large 4 'le Chonk,' Trained on 4,000 Grace Blackwell GPUs