Mistral Unveils Mistral Large 4 Preview, Pitching Trillion-Parameter Scale and Cybersecurity Capabilities
Mistral has launched a preview of Mistral Large 4, a European-trained 1-trillion-parameter model optimized for software vulnerability defense and enterprise workflows.
European AI startup Mistral has released a public preview of Mistral Large 4 (ML4), its largest artificial intelligence system to date. Nicknamed "le Chonk," the natively multimodal model contains one trillion parameters and is currently accessible via API on Mistral Studio, with open model weights scheduled for release at the end of October.
Mistral claims ML4 is the most capable open-weight model produced in either Europe or the United States across aggregated benchmarks, highlighting strong results in software vulnerability patching, manufacturing, finance, and visual grounding.
Cybersecurity as a Key Differentiator
Mistral is heavily positioning ML4 around enterprise IT security and cyber defense. In the Artificial Analysis Cyber Index, ML4 ranks among the top five models globally and leads non-Chinese open-weight models by a significant margin.
In a benchmark evaluating a model's ability to reproduce and subsequently patch real vulnerabilities in open-source software, ML4 scored 82 percent—the highest of any model tested. Mistral noted that leading closed models like Claude Opus 5.5 and GPT-6 Astra scored near zero on this test primarily because their safety filters refused to process the prompts.
Mistral argued that defending software requires verifying vulnerabilities firsthand, warning that restrictive safety filters prevent legitimate security work while malicious actors can bypass those guardrails through jailbreaks. ML4 is designed to deploy in private clouds or on-premises to eliminate operational risks during incident responses.
At the same time, Mistral highlighted that ML4 maintained the highest refusal rate among open models when faced with overtly malicious cyber prompts across JailbreakBench, StrongREJECT, and AgentHarm. The model also blocked 93.3 percent of attacks on Lakera's B3 AI Security Benchmark.
Benchmark Performance Across Code, Vision, and General Reasoning
On the Artificial Analysis Intelligence Index—which combines ten cross-domain benchmarks—ML4 scored 38 points. This marks a substantial upgrade over Mistral Large 3 (9 points) and Mistral Medium 3.5 (14 points), as well as Z.ai’s GLM-5.2. However, ML4 still trails top-tier closed models, with Claude Opus 5.5 (Max) leading the index at 58 points.
In software development benchmarks, ML4 scored 49.8 percent on the Artificial Analysis Coding Agent Index, finishing ahead of Deepseek V4 Pro and Qwen3.8 Max. In a blind code-quality evaluation conducted by Surge AI, human annotators placed ML4 second among five models with a score of 3.74 out of 5, trailing Claude Opus 5 (4.22) but outperforming GLM-5.3 and Kimi K3.
For visual processing, ML4 scored 42 percent on the Dense 200 benchmark, narrowly topping GPT-6 Astra’s 41 percent. Mistral states the model can examine documents, charts, CAD files, and gigapixel satellite imagery. On business workflows, evaluations by Vals.ai indicated that ML4 surpassed GPT-6 Astra in legal and financial domains.
Architecture and Sovereign Infrastructure
According to Mistral's technical documentation, ML4 utilizes a fine-grained mixture-of-experts (MoE) architecture with 1.05 trillion total parameters, of which 49 billion are active per token. It integrates a 1.6-billion-parameter vision encoder and features a context window spanning one million tokens.
The model was trained from scratch in Mistral's European data centers using 3,800 Nvidia Grace Blackwell GPUs, with training corpora encompassing over 160 languages. Ongoing reinforcement learning (RL) post-training utilizes approximately 3,000 GPUs, generating roughly 33 billion tokens daily. Mistral noted that the RL training phase has not yet plateaued and will yield further gains before the final weights release.
The massive compute push follows Mistral's recent 3 billion euro Series D funding round, alongside an $830 million loan secured in March for a data center near Paris, targeted at reaching 200 megawatts of European compute capacity by late 2027. Mistral CEO Arthur Mensch previously cautioned European officials against relying entirely on US-based AI models for sensitive defense and military code audits.
During the preview period on Mistral Studio, API usage is priced at $0.68 per million input tokens, $2.09 per million output tokens, and $0.07 per million cached input tokens. Documentation also outlines a standard tier priced at double those rates ($1.36 input, $4.18 output, and $0.14 cached inputs). Full technical reports, licensing details, and model weights will be published at the end of October.
