Mistral ships its trillion-parameter model as five legal challenges circle its data center
Mistral Large 4, unveiled in Abu Dhabi on Tuesday, is the first product evidence behind the €21 billion valuation struck in September. The pitch is efficiency and sovereignty; the benchmarks are unpublished, rivals sell tasks for pennies, and the data center under it all is contested in a Paris court.
Vincent Jiang · 3 min read
Mistral AI released Mistral Large 4 on Tuesday, a one-trillion-parameter multimodal model nicknamed Le Chonk, announced by chief executive Arthur Mensch at a conference in Abu Dhabi. It is available now only behind a public guardrail endpoint; the open weights follow in about three weeks, once safety testing ends 1. After a summer spent hosting cheaper Chinese models, the lab is claiming its product proves it is still a frontier lab, not an inference shop 1.
The efficiency claim is unaudited
Mistral says ML4 was trained entirely on its own compute with 4,000 Nvidia GPUs, "two to three times less than our Chinese competitors, and significantly less than the closed source competitors," according to VP Science Pierre Stock. That is a company-claimed figure; no independent count exists 1. Mensch told the Abu Dhabi audience the model is "above the Chinese models on certain aspects, including cyber," naming no Chinese model and no measure. Benchmark results are still pending 2. For now, the only measured thing about Le Chonk is its size.
Rivals sell a task for pennies
The floor under that premium keeps dropping. Xiaomi's MiMo-V2.6-Pro, a 1.02-trillion-parameter open-weight model, tops Artificial Analysis' open-weight index at 46 and costs $0.13 per task; OpenAI's GPT-6 Luna runs $0.07 34. Anthropic's Claude Opus 5.5 still leads the full index at 58 4. Against open weights at $0.13, Mistral's sellable advantage is not intelligence per dollar. It is auditability, European hosting and cybersecurity tuning, the wedge Mensch is selling in the Gulf and Asia: "There's an enormous desire and an enormous need for alternative technology suppliers" 2.
Rivals price a task between $0.07 and $1.99; Mistral's flagship is unpriced
Data
| Value | |
|---|---|
| GPT-5.6 Sol | $1.99/task |
| GPT-6 Sol | $1.06/task |
| MiMo-V2.6-Pro (Xiaomi) | $0.13/task |
| GPT-6 Luna | $0.07/task |
The data center underneath is in court
Le Canard enchaîné reported on September 30 that Mistral's mega data center, Campus IA at Fouju in Seine-et-Marne, faces five legal challenges; four recours were filed that day before the Paris administrative appeals court to annul the prefectural permits. The project, roughly €50 billion, covers 70 hectares of farmland 40 kilometres south of Paris, is due to power on in 2027, and is expected to consume the equivalent of a nuclear reactor 5. These details rest on one French investigative outlet and its subscriber-only report; English-language coverage has yet to appear. Meanwhile the sovereignty offering keeps selling: TotalEnergies has handed Mistral a century of geological research to apply to drilling, per the same outlet 5, and Mensch is courting Gulf customers 2.
ASML and Samsung get their product test
The Series D round of $3.5 billion values Mistral past $24 billion, nearly double the $13.7 billion Series C struck a year to the day earlier, with Samsung leading and ASML, Nvidia, EQT, BlackRock and Advent in the register 67. ML4's stated optimized use cases, cybersecurity, finance and chip design, sit exactly where its two lead investors live: ASML led the Series C, Samsung the Series D 1. Benchmarks, the weight drop in three weeks and four pending court filings, not a stage in Abu Dhabi, will decide whether €21 billion bought Europe's frontier lab or its most expensive sovereignty pitch.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
