Mistral opens a public preview of Mistral Large 4, a 1.05T-parameter open-weight model trained in Europe

Mistral put Mistral Large 4 (ML4, nicknamed “le Chonk”) into public preview on October 6, 2026, calling it its largest model yet and Europe’s first trillion-parameter open-weight model. The preview API is live in Mistral Studio today; the weights are promised by the end of the month, and the license is not yet published.

What happened

Mistral’s announcement and model docs describe a roughly 1T-parameter mixture-of-experts model in public preview under the label “open-weight” — not “open-source.”

• Public preview now. Available through the Mistral Studio/API; weights “drop end of this month.” No distribution channel named. [1]
• Trained in Europe. From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, which also serve the preview. [1]
• Preview red-teaming. Cybersecurity leaders, vetted partners, and state authorities get a build with reduced moderation and expanded cyber capabilities. [1]
• No license yet. The terms that will govern self-deployment are unpublished. [1][2]

Why it matters

Mistral’s pitch is sovereignty: EU-hosted inference under European law, plus self-deployable weights. It trained on 160+ languages, including every official EU language, and is natively multimodal on input. [1][2]

That positions ML4 as the Western counterweight to Chinese open models in the trillion-parameter tier it names — GLM-5.3, Kimi K3, and DeepSeek V4 Pro. For EU public-sector, finance, and legal teams avoiding US-hosted APIs, an EU-run endpoint plus downloadable weights is the concrete pitch.

The model in numbers

Parameter discrepancy. The announcement says “1 trillion-parameter” and “49 billion active”; the docs say 1.05T total and 52B active. This article uses the docs figures. [1][2]

Pricing and access

Docs show two price columns per million tokens — the preview rate and the list rate. Mistral states no end date for the preview discount. [2]

Access is via Mistral Studio and the API, with a Mistral-operated European deployment plus other regions. [1] DataNorth lists GLM-5.3 near $1.40/$4.40 and Kimi K3 near $3.00/$15.00 per million tokens. [6]

What Mistral claims it can do

Every figure below is Mistral’s own evaluation unless an independent evaluator is named.

• Cybersecurity. Top five globally on the Artificial Analysis Cyber Index and, per Mistral, ahead of open-weight models built outside China; 93% of 40 Cybench challenges; 82% on a vulnerability-reproduction test where Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task. [1]
• Agentic coding. 61.7% DeepSWE v1.1, 59.4% SWE-Atlas-QnA, 28.3% Terminal-Bench 4, and a 49.8% combined Coding Agent Index that Mistral says leads DeepSeek V4 Pro 0813 and Qwen3.8 Max. [1]
• Knowledge work. automationBench 59.9%; AA-Briefcase 1,393 Elo; vals.ai legal and finance evaluations Mistral says exceed GPT-6 Astra; and, per Mistral, Harvey’s Legal Agent benchmark ahead of all open-source models. [1]
• Multimodal. 42% on Dense 200 visual grounding, versus 41% for GPT-6-Astra, per Mistral. [1]
• Safety. 93.3% attack resistance on Lakera’s B3 benchmark and KORA score 1.691, plus, Mistral reports, a higher refusal rate on harmful cyber prompts than all open-weight models. [1]
• Independent check. Artificial Analysis scores the preview 38 on its Intelligence Index — above the median but not the frontier — and currently labels it proprietary because weights are unavailable; Brussels Signal reports seven Chinese open-weight models rank above it. [4][5]

Risks and caveats

• License pending. “Open-weight” is Mistral’s term; the actual terms, including commercial-use limits, are unpublished. Don’t plan self-hosting around assumptions.
• Preview pricing has no stated sunset, so the $0.68/$2.09 rate is not a guaranteed steady state.
• Benchmarks are self-reported. Except for Artificial Analysis and the named third-party evaluators, every score is Mistral’s, including the Cybench refusal comparison. “Outperforms any open-weight model developed in the US or Europe” and the “€3 billion Series D — largest equity round ever raised by a European tech company” are Mistral statements this article did not independently confirm.
• Parameter and context numbers differ across sources. Mistral’s announcement and docs disagree on parameter counts; Artificial Analysis lists a ~520K context for the served preview versus Mistral’s 1M claim.
• Dual-use cyber positioning. Mistral frames reduced-moderation access as defensive research; that framing, not an independent audit, is the basis for the claim.

What to watch

• Weight release and license text, promised by end of October 2026 — the point at which “open-weight” becomes checkable.
• Independent evaluations, including updated Artificial Analysis runs and any published Cyber Index detail.
• EU deployment evidence, since sovereignty claims depend on where inference actually runs and under which terms.

Sources