Mistral, the French AI startup that has styled itself as Europe's answer to Silicon Valley dominance, announced Mistral Large 4 on October 6, committing to release the weights of its 1-trillion-parameter model by month's end. The company said it trained the model (internally nicknamed "Le Chonk") from scratch on approximately 4,000 NVIDIA Grace Blackwell GPUs housed in European datacenters, with final safety tuning set to wrap by roughly October 27, according to Axios.
The release caps a three-day sequence in which two Western labs unveiled open-weight models crossing 500 billion parameters—Reflection AI debuted its Beam model on October 5, then Mistral followed a day later—both serving as direct responses to the Chinese releases that dominated headlines this summer. Moonshot AI's Kimi K3, at 2.8 trillion parameters, arrived with full weights in late July. Zhipu's GLM-5.3 launched in mid-August with coding benchmarks that caught attention across the industry.
"I don't want to live in the future in which any oligopoly controls closed access to this type of intelligence," Pierre Stock, Mistral's vice president of science, said in an October 6 interview with Axios. It's a blunt framing of what has become a central tension in AI development: whether frontier models should remain in the hands of a few proprietary labs or be released more widely, despite the risks.
China Moved First, and Bigger
The open-weights landscape shifted over the summer when Chinese startups began shipping models that dwarfed what Western labs had made publicly available. Kimi K3's 2.8 trillion parameters represented the largest openly accessible model at the time, according to VentureBeat and the South China Morning Post. GLM-5.3 followed three weeks later with strong coding and agentic performance, along with a phased plan to release its weights.
Meta's Llama 3.1, at 405 billion parameters, had set the Western high-water mark earlier this year. But no U.S. or European lab had crossed the trillion-parameter threshold with an open-weight commitment until October.
Reflection AI's Beam—a 501-billion-parameter mixture-of-experts model with 23 billion active parameters—appeared on October 5 with a promise to release weights later in the month, MarkTechPost reported. Mistral's announcement came a day later, describing Le Chonk as a 1.05-trillion-parameter MoE architecture with 52 billion active per token. The company said it trained the model over two months on approximately 4,000 Grace Blackwell GPUs, though Mistral's blog post cited 3,800 GPUs. The discrepancy wasn't addressed in public materials.
Stock acknowledged to Axios that the model "is not there yet on the frontier" compared to leading closed offerings from OpenAI or Anthropic. Guillaume Lample, Mistral's chief scientist, offered a more optimistic assessment in a VentureBet interview the same day, calling it "at the frontier of open weight models." Mistral's technical documentation lists a 1-million-token context window, multimodal input through a 1.6-billion-parameter vision encoder, and text-only output.
A European Infrastructure Bet
Mistral trained and now serves Le Chonk entirely on NVIDIA hardware in company-operated datacenters within the European Union. The setup is central to the startup's pitch around sovereign AI, the notion that European organizations should be able to run advanced models on European soil without relying on U.S. cloud providers.
The company has said it is targeting up to 1 gigawatt of EU compute capacity by 2030, according to an August blog post that announced regional inference endpoints and plans to host third-party open models. Zhipu's GLM-5.2 was named as the first outside model Mistral would serve on its infrastructure.
The training run represents one of the first large-scale European deployments of NVIDIA's latest GPU generation. NVIDIA communications have referenced multi-site Blackwell rollouts in Europe, with explicit mentions of collaboration in France. Mistral separately announced a partnership with Digital Realty in late May to expand localized AI services.
Funding for the infrastructure build came in part from a €3 billion Series D that closed on September 8, led by Samsung Electronics alongside EQT-managed Scaleup Europe Fund and PSG Equity. The round valued Mistral above €21 billion post-money, TechCrunch reported, making it the largest equity raise in European tech history. That's an extraordinary sum for a startup roughly two years old, and it reflects both the capital intensity of AI development and the geopolitical stakes Europe sees in having a credible alternative to American and Chinese labs.
Microsoft expanded its strategic partnership with Mistral on July 21, integrating the startup's models across Azure AI Foundry, Copilot Studio, and Azure Local. The deal explicitly targets regulated industries and sovereign deployments, according to a Microsoft press release. Airbus licensed Mistral's full suite in May with options for on-premises or trusted-cloud deployment. Cloudera integrated Mistral's models with its hybrid data platform (including air-gapped environments) in September.
Mistral's customer page lists companies including Airbus, ASML, and HSBC among its clients, with vendor claims of serving "Europe's top three banks and a third of the world's largest banks," though these promotional statements are undated and their current scope has not been independently verified.
Enterprise Adoption Is Shifting
Open-weight models now account for 34 percent of enterprise AI token usage, up from 23 percent a year earlier, according to a September survey of 200 organizations by Constellation Research and Enterprise Technology Research. Forty-two percent of respondents reported having open-weight models in production, with another 43 percent piloting them. A separate McKinsey survey fielded earlier this year found that 44 percent of organizations reported AI scaling enterprise-wide, up from 38 percent year-over-year, though the exact publication timeline of that data remains somewhat unclear.
Mistral launched Le Chonk's public API preview at promotional pricing: $0.68 per million input tokens (down from a standard $1.36), $0.07 for cached input, and $2.09 per million output tokens (down from $4.18). The pricing positions the model against premium closed offerings for cost-sensitive workloads. Stock told Axios that enterprise customers want "control, continuity, customization, cost"—four factors open weights can address more directly than closed APIs.
The company reported roughly 1,200 employees (about 300 researchers) and expects to reach approximately €1 billion in revenue in 2026, Le Monde wrote in September, citing company statements. Those figures are press-reported and have not been independently audited.
CEO Arthur Mensch told Le Monde that "AI is software. It can be controlled," pushing back against what he called "fear marketing" by closed-model vendors. He added that Mistral had raised €6 billion over three years, a figure that includes the recent Series D.
Technical Performance and Benchmarks
Le Chonk is structured as a granular mixture-of-experts model with 1.05 trillion total parameters and 52 billion active per token, according to Mistral's documentation. Some press accounts cite 49 billion active. Hardware analysis site QDNA noted on October 7 that the 49-billion figure may represent active parameters excluding embeddings and output layers, which would push the total to roughly 52 billion. The model accepts text and images as input through a 1.6-billion-parameter vision encoder but outputs text only, VentureBeat confirmed.
Mistral published a set of benchmark results on October 6 showing the model at 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, 28.3 percent on Terminal-Bench 4.0, and a composite "Coding Agent Index" of 49.8 percent, according to the company's blog. On agentic workflows the company cited 59.9 percent on AutomationBench-AA, a 657-task variant developed by Zapier and Artificial Analysis. For finance the model scored 67 percent on Finch/FinWorkBench, a benchmark reportedly published in ACL Findings earlier this year.
On legal tasks the company said Le Chonk achieved a 15 percent pass rate on the Harvey Legal Agent benchmark, compared to roughly 12.9 percent for Kimi K3 and 8.3 percent for GLM-5.3 on the Vals.ai leaderboard. Mistral's preview entry did not appear on that public board as of October 8, however.
VentureBeat noted on October 6 that several of Mistral's comparative scores for competing models are vendor-aggregated and that independent leaderboards show different absolute ranks, with Le Chonk not yet widely listed as of the article's publication. Mistral also claimed a 93.3 percent score on Lakera's b³ attack-resistance framework, calling it the highest among competitors, though third-party reproductions were pending at the time of the announcement.
The benchmarks matter less than real-world performance, and that's still being tested. Early access partners received a less-moderated variant during the preview window for red-teaming, according to Mistral's blog, which suggests the company is taking safety tuning seriously even as it prepares for a public release.
Regulation and the Sovereignty Angle
The European Union's AI Act began phased enforcement on August 2. The regulation presumes that models trained with more than 10^25 floating-point operations carry systemic risk and must notify the European Commission within two weeks of crossing the threshold, according to Regulation 2024/1689 and official guidelines. The guidelines list 10^23 FLOP as an indicative threshold for general-purpose AI identification and note the AI Office may designate models below 10^25 FLOP based on supplementary criteria.
Mistral has not publicly disclosed whether Le Chonk meets the systemic-risk compute threshold or filed a notification.
Mistral said it will release Le Chonk under a "custom Mistral license," VentureBeat reported, though the exact terms were not posted as of October 8. The company has historically used licenses that allow commercial use with some restrictions on competitive offerings. Stock told Axios the October 27 target follows additional reinforcement learning and safety tuning.
Policy frameworks for open-weight models remain inconsistent. The UK has published an approach to frontier model evaluations focused on cybersecurity and misuse risks. U.S. policy remains fragmented across states with no unified federal framework, Axios reported on October 6. The European Commission advanced a sovereign cloud procurement initiative in April that referenced consortia including Proximus, S3NS, and partners involving Mistral for AI services, according to a Commission press release.
What Happens Next
The arrival of multiple Western models above 500 billion parameters in a single week marks the first sustained response to China's open-weight push. But the competitive picture remains fluid, perhaps more so than the announcements suggest.
Chinese labs moved first and larger. Kimi K3's 2.8 trillion parameters still exceeds any Western release, and its weights shipped with fewer restrictions. Zhipu's GLM-5.3 topped several public leaderboards in August with coding and agentic scores that drew attention, the South China Morning Post reported. Mistral's own benchmarks acknowledge it trails leading closed models on raw capability.
The shift to open weights at trillion-parameter scale changes enterprise economics and control dynamics. Constellation Research found in September that open-weight models now represent more than a third of production AI workloads, driven by cost, customization, and data-residency requirements. Mistral's European datacenter strategy and partnerships with Microsoft, Airbus, and Cloudera position the company to capture regulated-sector demand that closed APIs struggle to address, particularly in finance, aerospace, and government.
Stock's comment about avoiding "oligopoly control" of intelligence reflects a broader industry debate over who gets to train, deploy, and modify frontier models. The open-weights movement (now spanning Meta, Mistral, Reflection, Moonshot, Zhipu, and others) argues that concentrating AI capability in a handful of closed vendors creates dependency risks. Critics cite safety and misuse concerns that delayed releases and licensed access aim to mitigate.
Whether Le Chonk's weights appear on schedule will test Mistral's execution and signal how seriously Western labs take the competitive threat from China. Independent benchmark reproductions and real-world performance in coding, agentic workflows, and vertical tasks will determine whether the model closes the gap Stock acknowledged or simply narrows it.
For now, the race is on. The infrastructure is European, the weights are promised by month-end, and the question of who controls frontier AI remains very much unsettled.
