On Thursday, Mistral published the full weights of Mistral Large 4 — a trillion-parameter model anyone can now download and run on their own servers — three weeks after it debuted as an API-only preview, according to TechCrunch.

The specs are the headline and the small print at once. Large 4, nicknamed "Le Chonk," is a mixture-of-experts model with about 1 trillion total parameters and roughly 49 billion active for each token, a 1 million-token context window, and native text and image input, per the Hugging Face launch post. It was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters, with day-one support in Transformers, vLLM, and llama.cpp — and the 2-bit quant reportedly runs on a single Apple Silicon Ultra with 512GB of unified memory. Three extra weeks of reinforcement learning between the October 6 preview and Thursday's release lifted its Terminal-Bench score from 28.3% to 41%.

The pitch is cybersecurity. Mistral says the model scores 51% on the Artificial Analysis Cyber Index and refused 96.6% of malicious cyber requests in its tests; it reports 82% on a task that asks a model to reproduce a real vulnerability and then patch it — a task where, llms.blog notes, several closed frontier models score near zero because they refuse the work entirely. Booz Allen, CrowdStrike, Elastic, and Europe's cybersecurity agency ENISA said they've been testing it, with Cisco and Cloudflare backing the open approach, per Skimwire.

"As a business company or as a government, you're able to have access to frontier intelligence to defend yourselves against cyberattacks," Mistral VP of science Pierre Stock told TechCrunch. "Access that can't be cut off by definition because the model is open-weighted."

That sentence is also the risk disclosure. Open weights can't be recalled, and Mistral is unusually blunt about it: the company says it saw "a potential for uplift to conduct real attacks" during the preview, and its more aggressive pen-testing variant stays restricted to vetted partners like Orange Cyberdefense, under contract. The license adds its own carve-out — the Mistral Ecosystem Open-Weight License is an MIT derivative that permits use, modification, and commercialization, except that anyone selling the model as a hosted service needs a separate commercial license. You can own the weights; you can't rent them back to everyone else.

The contrast is the story. Google's Gemini 4 Carbon, announced Wednesday, is a frontier cybersecurity model available only to vetted defenders. Mistral went the opposite way, and CEO Arthur Mensch's framing is ideological: AI should "be a tool for empowerment, not dominance." The benchmark tables, though, say Europe is still chasing — Large 4 is the top Western open-weight model on the Artificial Analysis Intelligence Index at 26, but only fifth overall, behind the Chinese open models.

The honest takeaway: the open-model race has become a cyber race, and Mistral just handed its heaviest weapon to everyone — defenders and attackers alike. The next weeks of independent testing will decide whether the guardrails survive contact with the download button.

— via TechCrunch