Mistral Large 4, aka “Le Chonk”: the French heavyweight against GPT-6 Astra and Claude

Science without conscience is only ruin of the soul. An AI model named like an obese meme cat, but armed with 1 trillion parameters. Mistral unveiled Large 4 on October 6, unofficially “ML4”, very officially “the Chonk” according to its own post. The Parisian start-up, which raised 3 billion euros in September, says it is at the state of the art among open models and claims a short lead over GPT-6 Astra in finance, with Claude remaining ahead. The figures come from the publisher, so handle with a pinch of salt.

Large 4: what Mistral announces

According to Mistral, ML4 is a multimodal model with a “mixture of experts” architecture (a set of subnetworks of which only part is activated for each request): 1,000 billion parameters in total, including 49 billion active. Trained from scratch on approximately 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers, in about two months according to VentureBeat, it covers more than 160 languages, including all official languages ​​of the European Union. A preview version is available via the API. Model weights are due to follow at the end of October.

Sovereignty is the main selling point. Mistral speaks of a European deployment operated end-to-end, “independently of other digital service providers and under European law”. For a bank or a fund subject to the GDPR, the argument weighs. VentureBeat specifies that the weights will be released on October 27 under in-house license, which Mistral did not detail in its announcement.

“This is Mistral Large 4, aka Le Chonk. 1,000 billion parameters, natively multimodal, 49 billion active. It is the best open weight model from the United States or Europe on aggregated benchmarks. It is state of the art for critical uses, including cyberdefense, industry and finance […]”

Mistral AI, on X

Benchmarks: ahead of GPT-6 Astra in finance, behind Claude elsewhere

On finance, Mistral claims that ML4 outperforms GPT-6 Astra on legal and financial tasks. According to Decrypt, the Finance Agent v2 test gives 54.7 to Large 4 compared to 53.5 to GPT-6 Astra. Claude Opus 5.5 is however ahead, at 58.6. A difference of 1.2 points on a test provided by the publisher does not decide between anyone.

Elsewhere, the competition remains ahead. On the human evaluation of Surge AI code quality, ML4 obtains 3.74 out of 5 compared to 4.22 for Claude Opus 5. On AutomationBench (657 business software tasks), it scores 59.9 compared to 71.8 for Claude Sonnet 5.5. In DeepSWE 1.1 encoding, it scores around 62 compared to 68 for Kimi K3.

Price and finance: the portfolio argument

The price displayed is $1.36 per million tokens for entry and $4.18 for exit. According to Decrypt, that's about a third of Claude Opus 5.5's entry price and a fifth of its release price. For massive uses, such as analyzing reports or processing credit files, the invoice counts as much as the score.

The publisher also puts forward the possibility of deployment in private cloud or on site, a sensitive point for financial data. The results remain to be confirmed by independent tests after the release of the weights. The start-up, valued at more than 21 billion euros after its September turn, is playing its credibility at this meeting.