Mistral shipped Large 4 on October 6, 2026, and called it an open-weight model. As of today, October 12, you can’t download it. What exists is a preview API on Mistral Studio, a promise that “weights drop end of this month,” and a model card that never names a license. That distinction matters if you plan to build on it.
What’s actually shipped#
According to Mistral’s announcement, Large 4 is a natively multimodal mixture-of-experts model with about 1 trillion parameters, 52 billion active, merging instruction and reasoning modes. Outlets disagree on the exact size (MarkTechPost cites 1.05T), so treat the numbers as approximate until a checkpoint exists to count.
Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. That sits well below Claude Sonnet 5.5’s $2/$10 but far above DeepSeek-class pricing. Mistral also stresses European hosting: trained and served in its own EU datacenters, with a private-cloud or on-prem cyber offering for vetted partners.
The benchmark picture#
Every headline number comes from Mistral:
- DeepSWE v1.1: 61.7%
- Terminal-Bench 4.0: 28.3%
- SWE-Atlas-QnA: 59.4%
- Cybench: 93%; vulnerability reproduce-and-patch: 82%
- Blind human coding rating: 3.74/5, second behind Claude Opus 5 (4.22)
The agentic coding runs were done privately, so nobody can reproduce them. The independent data is less flattering. Artificial Analysis measured Terminal-Bench 4.0 at 26.8% (close to Mistral’s figure, to be fair) but put Large 4 at 38 on its Intelligence Index, behind GLM-5.3 (45) and Kimi K3 (44), per secondary coverage. On the public DeepSWE board, closed models sit around 74% and the leading Chinese open models near 69%. Large 4 isn’t listed yet.
Credit where due: Mistral publishing its own human-eval loss to Opus 5 is more honest than most launch posts. But 28% on Terminal-Bench 4.0 against Opus 5.5’s self-reported 66.4% tells you where this model sits for autonomous work. It isn’t a Claude Code replacement; it’s a sovereign-hosting option.
The license is the real question#
Large 3 shipped under Apache 2.0. Large 4’s page says only “open-weight.” One outlet predicts a custom Mistral license; that is speculation, not a confirmed term. We’ve seen this movie recently: Kimi K3’s open-weights license surprise showed how much “open” can hide in the fine print. A custom license with revenue caps or use restrictions changes whether a model is a foundation or a liability.
What to do with this#
- Don’t plan around it yet. No weights, no license, no independent coding leaderboard entry. Re-evaluate after the end-of-October drop.
- Treat it as a data-residency play. If EU-hosted inference is a hard requirement, the preview API is worth a pilot on non-critical workloads.
- Keep your specs model-agnostic. This is the case for Spec-Driven Development: a well-written spec lets you swap Large 4, Sonnet 5.5, or Kimi K3 behind the same harness and measure the difference on your own repo instead of trusting launch tables.
- Keep the primary agent on Claude. For long-horizon autonomous work, the gap on Terminal-Bench 4.0 is not close.
I’ll revisit when the weights and license land. Until then, “open-weight” is a roadmap item, not a property.
Sources#
- Mistral: Large 4 announcement
- LLM Gateway: new models this week
- Let’s Data Science: weights this month, license blank
- CNBC: Mistral unveils new AI model (search snippet only; page not readable)