Mistral Large 4 Goes Open Weight — 47% Fewer Tokens, 96.6% Refusal Rate on Malicious Cyber

Mistral released the weights of Mistral Large 4 (Le Chonk) on 29 October, three weeks after the model went into public preview. The 1-trillion-parameter, natively multimodal MoE was trained from scratch on European infrastructure, and the headline pitch is not raw capability but efficiency: doing more work per token, and refusing the wrong things reliably.

The Coding Claim Is About Efficiency

On Cognition’s private FrontierCode 2.0 benchmark, Mistral Large 4 scores 19.7%, ahead of Beam 1.0 (2.9%), GLM-5.3 Flash (14.2%), Kimi K3 (17.3%) and GLM-5.3 (17.8%). On Terminal-Bench 4.0 it improved by 13 points to 41% over the preview while cutting output tokens by 30%.

On the Artificial Analysis Coding Agent Index it scores 48% — effectively tied with the leading open-weight model, GLM-5.3 — while using 47% fewer tokens per task. Three reasoning modes now let callers pick the tradeoff: Low for simple work, Medium for everyday tasks, and High (the default) for precision work. Mistral also reports 73.3% on FinWorkBench and 84.5% on BrowseComp, the latter up from a 75% SFT checkpoint.

Security Cuts Both Ways

Mistral scores 51% on the Artificial Analysis Cyber Index while averaging a 96.6% refusal rate on malicious cyber prompts — the highest Mistral measures among open models — and, in its words, still completes legitimate defensive work without unnecessary refusals. Agent attack resistance on the B3 benchmark rose to 97.1% from 93.3% in the preview.

The candid admission is more interesting than the numbers. Mistral says it monitored cyber-related use in live traffic, saw the model refuse black-box attacks against live targets, but also sees potential for uplift to conduct real attacks — so guardrails are being modified, and a specialized cyber variant goes to vetted design partners such as Orange Cyberdefense, restricted by contract to owned or authorized systems.

The endorsement list reads like a security-vendor billboard: NVIDIA (accelerated infrastructure), Elastic, CrowdStrike, Cisco, Cloudflare, Booz Allen, Fortinet, Rubrik and Zscaler, with data-residency and sovereignty the recurring pitch.

The caveat worth carrying: these are largely self-reported numbers, some on private benchmarks, and “open weight” says nothing about license terms or the hardware bill for hosting a trillion-parameter model. The interesting story, though, is the combination — a frontier-adjacent open model whose headline metric is tokens saved, and whose safety story is a dual-use posture the vendor is openly re-tuning.