Mistral Launches Trillion-Parameter Model, Challenging US and Chinese AI Rivals

Mistral AI, the Paris-based artificial intelligence company that has positioned itself as Europe's answer to OpenAI and Google, on Tuesday opened a public preview of Mistral Large 4, a trillion-parameter model the company says is the strongest open-weight system built outside China. The model, which Mistral's own staff have nicknamed "Le Chonk," is available now through the company's paid API, with the underlying weights scheduled for a wider release at the end of the month.

A trillion-parameter wager

Mistral Large 4 is a Mixture-of-Experts system with roughly 1.05 trillion total parameters, of which only about 49 billion are activated for any given task, according to the company's own release. That design lets a very large model run at a fraction of the computing cost of activating every parameter at once. The model is natively multimodal, handling text and images in the same request, and Mistral says it supports context windows as long as one million tokens and was trained across more than 160 languages, including every official language of the European Union.

The company says the model was trained from scratch over roughly two months using close to 4,000 Nvidia Grace Blackwell processors running in Mistral's own European data centers, a notably lean buildout by industry standards. "We trained this model on 4,000 Nvidia GPUs, two to three times less than our Chinese competitors, and significantly less than the closed source competitors," Pierre Stock, Mistral's vice president of science, said in remarks reported by TechCrunch.

The numbers Mistral is advertising

Access to the preview model runs through Mistral's developer platform at $1.36 per million input tokens and $4.18 per million output tokens, pricing listed in the company's own model documentation. On the benchmarks Mistral chose to publish, Large 4 scored 61.7 percent on the DeepSWE 1.1 coding-agent test and 93 percent on Cybench, a measure of a model's ability to find and patch software vulnerabilities, placing it, by the company's account, among the top five systems worldwide on an aggregate cybersecurity index and the strongest open-weight entrant outside China. The Next Web's review of the release noted additional claims, including a 67 percent score on a finance-focused benchmark called FinWorkBench, roughly tying the Chinese model DeepSeek-V4-Pro, and a markedly weaker 15 percent on legal-reasoning tasks, which Mistral and outside reviewers alike described as a weak point for every model tested, not just its own.

• Total parameters: roughly 1.05 trillion, Mixture-of-Experts architecture
• Active parameters per task: about 49 billion
• Training hardware: approximately 4,000 Nvidia Grace Blackwell GPUs, in-house European data centers
• Context window: up to 1 million tokens; input/output pricing of $1.36 and $4.18 per million tokens
• Open-weight release date: October 27, 2026, with a less-restricted version for vetted cyber-defense users

Built on a record European funding round

The launch lands a month after Mistral closed a €3 billion Series D round, which the company and investors called the largest single equity raise by a European technology firm. Samsung Electronics led the round alongside the EU-backed Scaleup Europe Fund, managed by EQT, and PSG Equity, with BlackRock, the Grand Duchy of Luxembourg and Advent joining as new investors and Nvidia, ASML, Salesforce Ventures and Andreessen Horowitz returning, according to Mistral's funding announcement. The round valued the four-year-old company at more than €21 billion, almost double the €11.7 billion valuation it carried after its 2025 round led by ASML, a detail confirmed independently by tech.eu's reporting on the deal. Chief executive Arthur Mensch has said the capital will go toward data centers and computing capacity that Mistral plans to largely own rather than rent, with the goal of controlling nearly all of its own compute within five years.

That buildout underpins what Mistral calls "sovereign" artificial intelligence: models and infrastructure that let a customer's data, workflows and institutional knowledge stay inside its own walls rather than passing through an American or Chinese cloud provider. The pitch has found an audience among regulated industries and governments wary of depending on a small number of foreign AI vendors. Mistral says it now serves more than 125 enterprise customers across 20 countries, including Airbus, the chipmaker ASML and the bank HSBC.

Who stands to be affected

The immediate audience for Large 4 is developers and enterprises already weighing which large language model to build on, particularly in Europe, where data-residency and regulatory concerns have made purely American or Chinese options a harder sell for banks, defense contractors and public agencies. Cybersecurity firms and government bodies are being offered a separate, less-restricted version of the model ahead of the general open-weight release, giving them earlier access to features Mistral is marketing for vulnerability research and incident response. Coverage of the launch in India framed the release chiefly as a challenge to the open-weight models coming out of China, which have set the pace on cost and availability over the past year, as well as to closed systems sold by OpenAI and Google.

What Mistral is saying about safeguards

Because Mistral intends to publish the model's weights, rather than keep them behind an API indefinitely, the company has had to explain how it will keep a freely downloadable trillion-parameter system from being repurposed by attackers for the same cyber capabilities it is marketing to defenders.

"We'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks," Pierre Stock, Mistral's vice president of science, said of the plan to vet early access to the model's cyber-oriented capabilities.

That framing leaves open how enforceable such a restriction can be once weights are downloadable by anyone, a tension that has followed every prior open-weight release with strong offensive-security capabilities, from Meta's Llama line to China's DeepSeek and GLM models.

What happens next

Mistral says it will publish Large 4's open weights on October 27, nearly three weeks after Tuesday's preview, giving outside researchers their first chance to independently verify the benchmark numbers the company has published so far. Until then, Large 4 exists only as a hosted preview, priced and distributed entirely on Mistral's own terms. The bigger test will come once independent labs and competitors run their own evaluations against the model and against the next releases expected from OpenAI, Google, Meta and Chinese labs including DeepSeek and Zhipu AI's GLM line, all of which have been locked in a rapid, monthly cadence of model releases through 2026. For Mistral, Large 4 is also a proof point for the compute strategy funded by its September round: whether a European company training on a few thousand GPUs, rather than the tens of thousands used by its best-funded rivals, can keep pace on performance while building the sovereign infrastructure its government and enterprise customers are paying for.