Paris, October 6. Mistral AI has pulled back the curtain — partially — on its new flagship model, Large 4. The headline numbers are the kind that make infrastructure engineers reach for a calculator: 1.05 trillion total parameters, with 52 billion active on any given forward pass, arranged in a natively multimodal mixture-of-experts architecture. Inside the company the model reportedly goes by “le Chonk,” which tells you most of what you need to know about its appetite for GPUs.
But note the operative word: preview. What launched today is API access through Mistral Studio, not the open-weights release the community is actually waiting for. That, Mistral says, arrives at the end of October. If you have followed Mistral’s release cadence before, you will treat that date as a plan rather than a promise — though the French tech press is already circling October 27 as the likely day. Until the weights land and the license is published (currently listed as “to be announced”), Large 4 is best understood as a very large API with an IOU attached.
A trillion parameters, 52 billion at a time
The mixture-of-experts design is doing the heavy lifting here. Instead of activating all 1.05 trillion parameters for every token, a router directs each token to a small subset of specialist sub-networks — 52 billion parameters’ worth, according to the model card. That is how you ship a trillion-parameter model without requiring a trillion parameters’ worth of compute per token. (One wrinkle worth noting honestly: launch-day coverage cited 49 billion active parameters, while the model card says 52 billion. Small discrepancy, but it is a reminder that preview numbers move.)
Multimodality is native rather than bolted on. A 1.6-billion-parameter vision encoder sits inside the architecture from the start, which matters for tasks like visual grounding — pointing at the right region of an image when asked — where retrofitted vision adapters tend to underperform. Training a MoE at this scale is its own engineering discipline, and the open-source world is racing to catch up: frameworks like Olmo-core 3, released this month by Ai2, are trying to make trillion-parameter MoE training reproducible outside the big labs.
There is a practical consequence to the MoE design that the parameter count obscures: serving Large 4 is a 52-billion-parameter problem, not a trillion-parameter one. The full 1.05T has to live somewhere — in memory or fast storage — but the compute per token scales with the active slice. That is the economic logic of the entire MoE wave: frontier-scale capacity with large-model-but-not-absurd inference bills. It is also why the open-weights release matters so much — a trillion-parameter MoE you can only access via API is a product; the same weights in the open are infrastructure.
The benchmark sheet, vendor edition
Mistral’s launch materials make the usual battery of claims, and they are worth reading with the usual grain of salt — every one of these is a vendor-reported number awaiting independent replication:
- DeepSWE v1.1 (agentic coding): 61.7%, a hair above GLM-5.3 at 61%.
- FinWorkBench (finance): 67%, matching DeepSeek-V4-Pro.
- Dense 200 (visual grounding): 42%, versus 41% for GPT-6 Astra — the payoff, presumably, of that native vision encoder.
- Artificial Analysis Cyber Index: a top-5 placement.
- Vals AI evaluations for legal and finance work, where Mistral claims Large 4 beats GPT-6 Astra.
The interesting pattern is where Mistral chose to compete: agentic coding, finance, legal, and vision. That is not a general-chatbot benchmark slate; it is an enterprise-workload benchmark slate, aimed squarely at the customers who pay per token.
Pricing, with a launch sale
Speaking of paying per token: Large 4 lists at $1.36 per million input tokens ($0.14 cached) and $4.18 per million output tokens, with an introductory sale cutting those to $0.68, $0.07, and $2.09 respectively. The aggressive caching discount is the tell — Mistral wants long-context, repeated-query enterprise workloads, the kind where prefix caching turns a good margin into a great one.
The sovereignty pitch
The less technical half of the announcement is geopolitical. Mistral says Large 4 was trained from scratch on 3,800 NVIDIA GPUs inside its own European datacenters, and the company leaned hard on the AI-sovereignty framing: European model, European infrastructure, European oversight. The model was red-teamed with security partners and state authorities before launch — a detail that reads as both genuine safety work and a signal to regulated-industry buyers. For more on why red-teaming claims deserve scrutiny, see our coverage of the “threat-model coverage gap” in AI safety evaluation.
The fine print on “open”
Mistral built its reputation on open weights, so the preview-first strategy deserves a fair reading and a skeptical one. The fair reading: releasing the API first lets the lab harden serving, gather safety data, and finalize the license before the weights go everywhere. The skeptical reading: every week the weights stay private is a week competitors can study the API’s behavior without reciprocating. Both can be true. What will settle it is the license text — “to be announced” covers everything from Apache-2.0 to research-only — and the calendar. End of October is close enough to verify soon.
What to watch
Three things will decide whether Large 4 matters beyond launch week. First, the weights: promised for end of October, license still unannounced. Second, independent benchmarks — vendor numbers are a starting point, not a verdict. Third, the multimodal story: if that 1.6B vision encoder delivers on visual grounding at scale, Large 4 could become the default open-weights choice for document-heavy enterprise work. Until then, it is a promising preview with a very French nickname — and an IOU the whole open-source community will be watching.
And then there is the competitive backdrop the benchmarks themselves reveal. Mistral chose to compare against GLM-5.3, DeepSeek-V4-Pro, and GPT-6 Astra — a tacit admission of who it considers the frontier. The gaps claimed are small (61.7% vs 61% on DeepSWE; 42% vs 41% on visual grounding), which is either honest reporting of a tight race or careful selection of benchmarks where the model looks best. Independent evaluation will tell us which. Either way, Large 4 confirms the industry’s direction of travel: the next round of open-weights competition will be fought between trillion-parameter MoEs, and the winner will be decided as much by license terms and serving economics as by benchmark points.
Featured image: the Mistral AI company logo (Mistral AI, via Wikimedia Commons — public domain).
