Skip to content
NewsModel release

Mistral opens the weights for its trillion-parameter Large 4 model, aimed at cybersecurity

· by Pondero Newsdesk

The short version

Mistral released open weights for Mistral Large 4 on October 29, 2026, a 1-trillion-parameter model it says scores 82% on a vulnerability-reproduction test where Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.

Mistral opens the weights for its trillion-parameter Large 4 model, aimed at cybersecurity

Mistral released open weights for Mistral Large 4 on October 29, 2026, and the number that stands out is not the model's size. It is that on a test asking a model to reproduce a real software vulnerability, Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task, while Large 4 scores 82%, according to TechCrunch.

What

Mistral previewed the model, nicknamed "Le Chonk," through a moderated API on October 6 and released its weights on Hugging Face and via its own API on October 29, per Mistral's launch post. Large 4 is a mixture-of-experts model with roughly 1 trillion total parameters and 52 billion active parameters, trained on 3,800 Nvidia Grace Blackwell GPUs inside Mistral's own European datacenters, and drawing on training data spanning more than 160 languages. Mistral positions it for cybersecurity, coding, agentic work, and multimodal tasks. On the Artificial Analysis Cyber Index's vulnerability-reproduction test, the company reports an 82% score, the highest of any model tested, plus a 93% result on Cybench, a 40-exercise benchmark drawn from security competitions. TechCrunch reports the model uses 47% fewer tokens per task than the Chinese open-weight model GLM-5.3 and that Mistral claims it ranks first among open models on Artificial Analysis' Coding Agent Index. Vice President of Science Pierre Stock told TechCrunch the open-weight release gives governments and enterprises "access to frontier intelligence to defend yourselves against cyberattacks" that "can't be cut off by definition because the model is open-weighted."

On coding specifically, Mistral reports a 61.7% score on DeepSWE v1.1 and a combined 49.8% on Artificial Analysis' Coding Agent Index, which it says puts Large 4 ahead of the open models DeepSeek V4 Pro and Qwen3.8 Max, per the launch post. A separate blind human evaluation Mistral ran with Surge AI, where professional annotators rated outputs on a 1 to 5 scale with model identities hidden, ranked Large 4's preview version second of five models tested at a score of 3.74, behind only Claude Opus 5 at 4.22 and ahead of Kimi K3 and two GLM versions. On multimodal tasks, Mistral says the model beat GPT-6 Astra on the Dense 200 visual-grounding benchmark, 42% to 41%.

Open weights remove the kill switch, not just the price tag

The 82% score matters less on its own than what it implies about closed-model refusals. Reproducing a vulnerability before patching it is standard incident-response work, and Mistral's framing is that safety filters built into closed models can block exactly that work mid-incident, a risk the company says is itself a security problem. An enterprise or government security team weighing Large 4 against a closed frontier model is therefore not just comparing benchmark scores; it is comparing whether the tool can be taken away from them, by a vendor policy change, an account suspension, or a service outage, in the middle of responding to an attack. Running the weights on-premises or on a sovereign private cloud, as Mistral is positioning it, removes that single point of failure, at the cost of the security and moderation layer a closed-model provider would otherwise maintain on the vendor's side.

Context and reactions

Mistral CEO Arthur Mensch framed the release as philosophical as well as technical, telling TechCrunch that AI should "be a tool for empowerment, not dominance," and that Large 4 achieves "state-of-the-art performance" at "a fraction of the compute and capital committed by closed labs." The release lands as open-weight labs and closed-model providers increasingly compete on different axes rather than the same benchmark leaderboard: Mistral measures itself against GLM-5.3, DeepSeek V4 Pro, Kimi K3, and Qwen3.8 Max on coding and agentic tasks, per its own benchmarks, while reserving the raw-capability comparison against Claude Opus 5.5 and GPT-6 Astra specifically for the cybersecurity-refusal point. Stock also addressed a separate concern TechCrunch raised about rogue AI agents, saying Mistral does not "want to fearmonger people" and calling AI security "a hard problem" but "a problem that is solvable."

Mistral says the same training run also produced the base for what it calls a coming generation of specialized models built on Large 4's architecture, and that the model uses the same training, customization, and reinforcement-learning environment the company already sells to enterprise customers through its Forge product, per the launch post. That overlap matters commercially: a bank or manufacturer already paying Mistral to fine-tune a smaller model on Forge can now point the same tooling at a frontier-scale base model instead of switching vendors for the capability jump.

What to watch next

Mistral is still running a separate, more permissive version of Large 4 with cybersecurity leaders, vetted partners, and government testers under reduced moderation, per its launch post; independent red-team results from that program, rather than Mistral's own benchmarks, will be the real test of the safety tradeoff. Watch too for whether Anthropic or OpenAI respond to the refusal-rate comparison, and whether another open-weight lab matches Large 4's cyber positioning rather than conceding that lane to Mistral and GLM-5.3.

Sources