Signals›Open-source AI
Open-source AIShipping

Mistral releases the weights for its trillion-parameter model, pitched at cyber defense.

Mistral Large 4 launched as an API preview on October 6. Three weeks later anyone can download it — and Mistral is making a security argument for why that matters.

ZN ZilverNet EditorialPublished · 5 min read · 3 sources Weights released Oct 29, 2026
Mistral's launch artwork: the words '#1 in Sovereign AI.' above orange and red stripes
Image credit: Mistral · Mistral's launch artwork. "#1 in Sovereign AI" is Mistral's slogan, not our assessment.
The short versionShipping
What it is
A 1-trillion-parameter, natively multimodal mixture-of-experts model with 52B active parameters, nicknamed "Le Chonk."
How far along
ResearchPrototypePilotShippingEverywhere
Shipping. Weights on Hugging Face since Oct 29; the API costs $0.70 per million input tokens and $2.10 per million output.
Why it matters
An open-weight model can run on your own servers, where nobody can withdraw it or change its price.
Still unproven
Performance claims are Mistral's. No license is named in Mistral's announcement or in TechCrunch's coverage — check the download page before production use.

From preview to download

Mistral announced Large 4 on October 6 as a public API preview, promising: "Weights drop end of this month." On October 29 it delivered, publishing the weights on Hugging Face alongside its API.

The model has 1 trillion total parameters with 52 billion active, and takes images, documents, charts and PDFs as input. Mistral's announcement says it was trained on 3,800 Nvidia Grace Blackwell GPUs. Pierre Stock, Mistral's VP of Science, told TechCrunch that is "two to three times less than our Chinese competitors" — a company claim.

Why cyber defense

Mistral is positioning the release around security. "Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally," the company said.

The argument for open weights here is control. "Access that can't be cut off by definition because the model is open-weighted," Stock told TechCrunch. He was also careful about tone: "We don't want to fearmonger people. We think it's a problem that is solvable."

The three-week gap was deliberate. At launch, Stock said the weights would follow "after safety testing is complete," and that Mistral would work "with trusted partners and governments" in the meantime.

AI should "be a tool for empowerment, not dominance." — Arthur Mensch, CEO, Mistral

The claims, labeled as claims

  • Performance comparable to GLM-5.3 while using "47 percent fewer tokens per task."
  • It "outranks all other open models" on agentic coding tasks on Artificial Analysis' Coding Agent Index.

Both come from Mistral. Independent results are still to come.

The bigger picture

Samsung led Mistral's Series D at a €21 billion valuation, according to TechCrunch. French President Emmanuel Macron has called the model "a third way in AI."

ℹ

Neither Mistral's announcement nor TechCrunch's report names a license. Before building on Large 4, check the terms on its Hugging Face page — see our note on why the license matters.

Update log
  • Published.