Mistral Ships Trillion-Parameter 'Large 4,' Pitches It as Europe's Answer on Security
Mistral AI released Mistral Large 4 on October 6, a one-trillion-parameter mixture-of-experts model with 49 billion active parameters that the French lab is positioning less as a general chatbot upgrade and more as a sovereign, security-focused alternative to the dominant US frontier models. A public preview is live now through Mistral's Studio API, with full model weights promised by the end of October pending security review.
A deliberate pitch on the one axis closed labs avoid
The headline numbers are in cybersecurity, a category Mistral says it chose precisely because competitors won't compete there. The company reports Large 4 solves 93% of Cybench security-competition challenges and scores 82% on a vulnerability reproduction-and-patching benchmark, both of which it describes as the highest reported results to date. Mistral says the model also ranks in the global top five on the Artificial Analysis Cyber Index.
The comparison Mistral draws is pointed: it says Claude Opus 5.5 and GPT-6 Astra refuse many of the same offensive-security tasks outright, where Large 4 completes them. 'Defending software often starts with proving that a flaw is real,' the company wrote in its announcement, framing the model's willingness to do exploit-reproduction work as a feature for security researchers rather than a liability.
The model we're announcing today is above Chinese models on certain aspects, including cyber.
That line, attributed to Mistral CEO Arthur Mensch, frames the release against Chinese open-weight competitors like DeepSeek V4 and Qwen as much as against OpenAI and Anthropic — a sign of how three-way the frontier-model race has become by late 2026.
Beyond security: coding, multimodal, and knowledge work
Mistral also published results on agentic coding and business-workflow benchmarks, claiming a Coding Agent Index score of 49.8%, ahead of DeepSeek V4 Pro and Qwen3.8 Max, alongside a 61.7% result on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.0. On visual grounding, the company says Large 4 edges out GPT-6 Astra (42% versus 41% on the Dense 200 benchmark), and it claims the model beats GPT-6 Astra on legal and financial reasoning tasks, including Harvey's Legal Agent benchmark for document-heavy knowledge work.
On safety, Mistral reports the model resists 93.3% of attacks on the B3 AI Security Benchmark and posts a 1.691 score on the KORA benchmark, close to that scale's 2.0 ceiling — an attempt to show the model's permissiveness on offensive-security tasks hasn't come at the cost of guardrails elsewhere.
Infrastructure and the European-sovereignty angle
Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs inside Mistral's own European data centers, with a reinforcement-learning pipeline the company says generates roughly 33 billion tokens a day. Mistral is emphasizing that European deployment is operated independently and under European law — a continuation of the sovereign-AI argument it has used since its earlier Large releases, now paired for the first time with a claim of outright benchmark leadership rather than parity. Pricing for the preview API is set at $1.36 per million input tokens and $4.18 per million output tokens.