AI Wire

Short briefings on models, tools, agents, and the rules that govern them.

English edition

Models

Mistral Large 4 Brings 1T Scale to Enterprise Agents

Mistral AI launched Mistral Large 4, an open-weight model with 1.05T parameters tailored for multi-step enterprise workflows and legal automation.

Open-weight European model leads open rivals in workplace automation benchmarks ahead of full weights release in late October.

French AI lab Mistral AI has announced Mistral Large 4, internally codenamed “le Chonk”, an open-weight model engineered to run autonomous agents across complex enterprise workflows. The official announcement arrived through the startup’s channels and technical documentation. Both sources detail systems capable of coordinating external software tools, querying corporate databases, and generating completed deliverables without continuous human prompting.

The Mistral Large 4 architecture

The system relies on a Mixture-of-Experts (MoE) design, an architecture that routes tokens through specialized sub-networks so only a fraction of total parameters runs during each step. While the entire architecture spans 1.05 trillion parameters, only 49 to 52 billion remain active per token. The release also integrates a dedicated 1.6-billion parameter vision encoder and native support for more than 160 languages.

Mistral designed this balance to deliver high analytical capacity alongside workable inference costs for production environments. Rather than operating purely as a conversational assistant, the model was calibrated to follow extensive instruction chains, execute calls across workplace applications, and output structured spreadsheets or documents directly.

In benchmark evaluations published by Mistral on AutomationBench, which tests 657 enterprise workflows integrated across Gmail, Google Sheets, Slack, and Salesforce, the model achieved a 59.9% task completion rate. That score placed it ahead of international open-weight competitors, including DeepSeek V4 Pro, MiMo-V2.6-Pro, and Kimi K3.

In legal tasks, the company reported the top ranking among open models on the Harvey Legal Agent Benchmark (LAB), an evaluation run by Harvey AI focused on contract review and statutory research. Because the underlying weights are not yet open for download, these numbers currently reflect reports from Mistral and its testing partners. Independent third-party confirmation will depend on open weights distribution in the coming weeks.

European hardware and data sovereignty

To train the model, Mistral deployed an internal cluster inside European data centers powered by 3,800 NVIDIA Grace Blackwell GPUs. The training phase took roughly two months, financed through the company’s 3-billion-euro Series D funding round.

The choice of European infrastructure serves as a practical pitch to enterprises and public agencies operating under European Union privacy regulations and the EU AI Act. Organizations hesitant to send confidential records to cloud APIs hosted in the United States gain an alternative that can run locally on dedicated servers or regional cloud providers.

API preview and weights release timeline

Mistral Large 4 is available immediately for testing in public preview through Mistral Studio. Commercial API access is priced at $1.36 per million input tokens and $4.18 per million output tokens.

The complete release of open weights for direct download and self-hosting is scheduled for late October 2026. Once those files become public, outside engineering teams and academic researchers will be able to audit model behavior, test local hardware requirements, and deploy the system inside isolated corporate networks.

Sources

Sources

  1. mistral.ai
  2. docs.mistral.ai
  3. x.com
  4. heise.de
  5. veja.abril.com.br

Markers

  • mistral-ai
  • mistral-large-4
  • llm
  • open-weights
  • ai-agents

Corrections and editorial policy