Mistral AI Takes on OpenAI, Anthropic With Its 1 Tn Parameter Model

Mistral AI has launched a public preview of Mistral Large 4, a 1 trillion-parameter multimodal model nicknamed ‘Le Chonk’, as the French AI company looks to compete with leading models across coding, agentic workflows, cybersecurity and multimodal tasks.

The company said the model is available through its preview API, with the model weights expected to be released by the end of October. 

Mistral Large 4, internally referred to as ML4, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs at Mistral’s European datacentres. The company said the same infrastructure is being used to serve the public preview. 

“ML4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters,” Mistral said in a statement, calling it its largest model to date. The company said it has been developed for coding, agentic workflows and multimodal understanding, with applications spanning cybersecurity, finance and legal work. 

Mistral Large 4 Targets Cybersecurity

Cybersecurity is one of the areas where Mistral is positioning the model against closed AI models. On the Artificial Analysis Cyber Index, Mistral said ML4 ranks among the top five models globally. It scored 82% on a test involving reproducing and patching a real vulnerability in open-source software, while scoring 93% on Cybench. 

Mistral said the model can be deployed on private cloud or on-premise infrastructure, giving organisations control over how the system is used in security operations.

The company also pointed to the model’s performance on security tasks that some closed models refuse to perform.“Several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse to perform the task,” Mistral said.

“Defending software often starts with proving that a flaw is real,” it added, arguing that refusal behaviour can limit legitimate vulnerability research and incident response. 

Coding and agentic workflows

ML4 scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, according to Mistral. Its combined Coding Agent Index score was 49.8%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen 3.8-Max.

In a blind evaluation with Surge AI, ML4 ranked second among five models for coding quality, scoring 3.74 out of 5. It trailed Claude Opus 5 at 4.22 and ranked ahead of Kimi K3, GLM-5.3 and GLM-5.2.

The model also scored 59.9% on AutomationBench, which evaluates 657 business workflows across applications such as Gmail, Google Sheets, Slack and Salesforce. Mistral said ML4 also reached 1,393 Elo on AA-Briefcase, a benchmark for long-horizon knowledge work. 

Multimodal capabilities

Mistral Large 4 can process images, documents, charts and natural scenes. The company said it can combine visual understanding with agentic capabilities for tasks such as analysing engineering drawings, retrieving evidence from PDFs and examining large satellite images.

On the Dense 200 visual-grounding benchmark, Mistral said ML4 scored 42%, compared with 41% for GPT-6 Astra. 

The model also targets scientific workloads. Mistral said ML4 is state of the art among open-weight models on SciCode-Verified and can perform tasks including data analysis, modelling and scientific simulations.

Built around open weights

Mistral is positioning the model around open weights and self-deployment, particularly for organisations that need control over AI systems. The company said ML4 will be available across multiple regions, including a European deployment operated by Mistral under European law. 

More than 160 languages were represented in a significant share of the model's training data, including all official languages of the European Union, according to Mistral. 

The company said ML4 was developed with enterprises across finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping and the public sector. Its training and customisation environment is based on the same system Mistral offers customers through Mistral Forge. 

Weights coming this month

Mistral said it will release the model weights by the end of October along with more details about its architecture, benchmarks and post-training methods. 

The company said the model is part of a roadmap funded by its €3 billion Series D, which it described as the largest equity round raised by a European technology company. Mistral is also scaling compute capacity at its European datacentres. 

For now, Mistral Large 4 is still undergoing reinforcement-learning training. The company said that at its current scale of around 3,000 GPUs, a single training run produces about 33 billion tokens a day, including around 16 billion trainable completion tokens after filtering and masking.