Mistral opens Large 4 preview, a 1.05 trillion-parameter model with open weights promised

Mistral has opened a public preview of Large 4, a natively multimodal model with 1.05 trillion total parameters and 52 billion active, and says the weights will be released. The company puts it this way: "Unofficially ML4, very officially: le Chonk."

The preview API is live on Mistral Studio, and Mistral says the weights "drop end of this month." Mistral's announcement and documentation carry an October 6, 2026 date.

What is in the model

Mistral's documentation describes a granular Mixture-of-Experts architecture with 52B active parameters, 1.05T total, and a 1.6B vision encoder. The context window is 1 million tokens. The docs list structured outputs, function calling, document Q&A, batching, agents and conversations, and built-in tools.

Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the preview is served from the same infrastructure. It also says a significant share of the training data was multilingual, spanning more than 160 languages, including every official EU language.

Pricing is $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens.

The announcement gives no architecture breakdown beyond the parameter counts: no layer count, no expert count, no routing details. It also does not say which license will govern the weights, or where they will be hosted.

VentureBeat, meanwhile, reports different numbers: 49 billion active parameters and 4,000 GPUs, plus a specific release date of October 27 after a roughly three-week testing period. Mistral's own figures are 52 billion and 3,800, and its date is "end of this month." When the company and the trade press disagree on the size of its own model, I take the company's numbers and keep a pencil handy.

The scorecard, read slowly

Mistral supplies a lot of benchmarks. Here they are, with what each one does and does not say.

• Coding agents: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4. The combined Coding Agent Index score of 49.8% puts it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, according to Mistral.
• Blind human eval: in a Surge AI coding-quality test on a 1-5 scale, ML4 Preview ranked second at 3.74, behind Claude Opus 5 (4.22) and ahead of GLM-5.3 (3.60), Kimi K3 (3.59) and GLM-5.2 (3.40). Second place sounds good until you notice the field was five models.
• Knowledge work: 59.9% on AutomationBench (657 business workflows across Gmail, Google Sheets, Slack and Salesforce) and 1,393 Elo on AA-Briefcase, ahead of DeepSeek V4 Pro.
• Vision: 42% on Dense 200 visual grounding against GPT-6-Astra's 41%. A one-point gap, and the announcement gives no error bars.
• Security: 93.3% of attacks resisted on Lakera's B3 benchmark, and 1.691 out of 2 on KORA.

The cyber results deserve the most care. Mistral says ML4 scores 82% on a test in the Artificial Analysis Cyber Index that asks a model to reproduce a real vulnerability in open-source software and then patch it, "the highest of any model." It also reports solving 93% of the challenges in Cybench, a set of 40 exercises drawn from security competitions.

Then comes the footnote that does the heavy lifting. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the same test "because they refuse to perform the task." Scores driven by refusals do not establish a capability gap. It is a fair thing to report, but it is a different claim from "our model is better at security." The announcement does not say how those models would score if they complied.

Mistral also says that before the weights ship it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities. The announcement does not say what the public weights will look like on that axis.

Most of these numbers come from Mistral's own post. Third-party names appear (Artificial Analysis, Surge AI, Lakera, vals.ai, which Mistral says found ML4 beat GPT-6-Astra on legal and financial tasks), but the post gives no methodology for the comparisons. I would like the eval harness more than the chart.

What to watch

Mistral calls ML4 the first milestone funded by its €3 billion Series D, which it describes as the largest equity round ever raised by a European technology company. That makes this release a test of whether the money produces a model people will actually run.

The real test arrives when the weights do. Watch for the license, since "open weights" covers a range of terms. Watch for the hosting location. And watch whether independent evaluators reproduce the DeepSWE and cyber numbers once outsiders can run the model themselves.

Until then, ML4 is a promising table of results with a trillion parameters behind it. Show me the eval, not the vibe.