Mistral Launches Public Preview of Mistral Large 4, a 1-Trillion-Parameter Open-Weight Model

Mistral has launched a public preview of Mistral Large 4 (ML4), a 1 trillion-parameter natively multimodal model with 49 billion active parameters. The model is currently accessible via the Mistral Studio API, with full open weights scheduled for release by the end of the month. ML4 is designed to push the frontier of open-weight performance, particularly in coding, agentic workflows, and multimodal understanding.

Performance and Benchmarks

ML4 demonstrates state-of-the-art performance among open models for critical enterprise workloads, including cybersecurity, finance, and law. On the Artificial Analysis Cyber Index, it scores 82% on a test requiring the reproduction and patching of real vulnerabilities, the highest score of any model. It also solves 93% of challenges in Cybench. In coding, ML4 achieves a combined Coding Agent Index score of 49.8%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. Specific coding benchmarks include:

• DeepSWE v1.1: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal-Bench 4: 28.3%

In agentic workflows, ML4 scores 59.9% on AutomationBench, which tests 657 business workflows across apps like Gmail and Salesforce. For knowledge work, it reaches 1,393 Elo on AA-Briefcase. In multimodal tasks, ML4 surpasses GPT-6-Astra on visual grounding in the Dense 200 benchmark (42% vs 41%).

Infrastructure and Sovereignty

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters. This infrastructure supports the model’s public preview and future deployments. Mistral emphasizes AI sovereignty, noting that ML4 will be available in a European deployment operated end-to-end by Mistral under European law, independent of other digital service providers. The model’s training data included more than 160 languages, covering every official language of the European Union.

Safety and RL Methodology

The model has saturated Mistral’s benchmarks for robustness to indirect prompt injections, resisting 93.3% of attacks on the B3 AI Security Benchmark. It also shows a higher refusal rate for malicious cyber requests than other open-source models, addressing concerns about safety filters blocking legitimate security research. Mistral’s reinforcement learning (RL) pipeline currently generates approximately 33 billion tokens per day on 3,000 GPUs, with 16 billion trainable completion tokens. This RL approach allows the model to adapt to increasing task difficulty and breadth.

Future Roadmap

ML4 serves as the foundation for a new generation of specialized Mistral models. The development is funded by Mistral’s €3 billion Series D, the largest equity round ever raised by a European technology company. Mistral plans to scale up compute capacity in its European datacenters, expecting further improvements in the coming weeks and months. Pricing for the preview API is $1.36 per million input tokens and $4.18 per million output tokens.