Mistral launches 1T-parameter Mistral Large 4

Mistral just dropped its biggest heavyweight yet, and it is leaning hard into enterprise sovereignty.

The French AI lab introduced Mistral Large 4, an open-weight multimodal flagship packing roughly 1 trillion total parameters with 52 billion active at inference. Unofficially dubbed ML4, the team officially refers to it as "le Chonk."

Developers can already take the model for a spin via a public preview API on Mistral Studio, while downloadable weights are scheduled to land before October wraps up.

Forged on European silicon

Mistral wants to offer an escape hatch from the binary choice of US closed clouds and Chinese open-source options. ML4 was trained from scratch in Europe on 3,800 NVIDIA Grace Blackwell GPUs. Pierre Stock, Mistral's science lead, noted the lab pulled this off with roughly 4,000 GPUs, significantly fewer than what major rivals typically burn.

Under the hood, the native vision and text model covers more than 160 languages, including every official European Union tongue. Backed by heavyweights like ASML and Samsung, Mistral is clearly targeting regulated sectors that want end-to-end sovereignty.

A cyber powerhouse that refuses to refuse

The most dramatic numbers land in cybersecurity. On the Artificial Analysis Cyber Index, ML4 scored 82% on reproducing and patching real-world software vulnerabilities. Top-tier closed systems like Claude Opus 5.5 and GPT-6 Astra scored near zero on that exact test because their safety filters refused to execute the tasks.

That difference is practical. Security analysts analyzing malware or reverse-engineering exploits cannot afford provider-level guardrails shutting them down mid-incident. Mistral is granting vetted partners and state agencies reduced moderation during the preview, setting up an open-weight release that organizations can run fully on-prem.

It also punches high elsewhere, edging out GPT-6 Astra on Dense 200 visual grounding (42% vs 41%) and beating open rivals on Harvey's Legal Agent benchmark. Still, outside observers point out it trails leading Chinese alternatives like Kimi K3 and DeepSeek V4.1 Flash on raw coding benchmarks, while costing more per benchmark run on third-party tracking.

Why it matters

We think Mistral played this very smart by prioritizing unneutered enterprise capability over purely chasing generic chatbot leaderboards.

For security engineers, legal teams, and defense contractors, ML4 is a compelling reason to pay attention. You get frontier-grade reasoning that you can genuinely inspect, tune, and host inside your own perimeter, without third-party guardrails breaking your defensive tooling.

For the wider market, this proves European labs can still build top-tier frontier systems with fractionally fewer chips. But Mistral now has to prove the economics: running a 1-trillion parameter MoE locally is no small feat, and competing against cheaper, hyper-efficient open models will test whether customers are truly willing to pay a premium for sovereign control.