Mistral AI Releases Mistral Large 4 Open-Weight Model

At a Glance
| Item | Value |
|---|---|
| Publisher | Mistral AI |
| Published | 2026-10-29 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Mistral AI has released the model weights for “Mistral Large 4 (Le Chonk)," a 1-trillion-parameter native multimodal model. Following three weeks of reinforcement learning after its public preview, it achieves significant functional improvements in coding, cybersecurity, and inference efficiency. It offers top-tier coding performance as an open-weight model and is provided as a foundation for sovereign AI that can be self-hosted on a company’s own infrastructure.
Specifications
- Number of parameters: 1-trillion-parameter
- Architecture: native multimodal (natively multimodal)
Performance
According to measurement results published by Mistral AI, Mistral Large 4 demonstrates top-tier results among open-weight models in coding, security, and general work tasks.
The main evaluation results shown in the model card and announcement are as follows:
| Benchmark / Evaluation Metric | Score / Measurement | Comparison and Details |
|---|---|---|
| Cognition – FrontierCode 2.0 | 19.7% | 1st place |
| Artificial Analysis Coding Agent Index | 48% | Equivalent to GLM-5.3 (47% reduction in average tokens used per task) |
| Terminal-Bench 4.0 | 41% | 1st place among open-weight models (13 points improvement from preview version, 30% reduction in output tokens) |
| Surge Code (human preference evaluation) | – | Outperformed open-weight models (GLM-5.3, Kimi K3, DeepSeek V4.1 Flash), ranking just behind Claude Opus 5.5 |
| Artificial Analysis Cyber Index | 51% | Vulnerability discovery, reproduction, and patching tasks |
| Refusal rate for harmful cyber requests | 96.6% | Highest among open models (average of JailbreakBench, StrongREJECT, AgentHarm) |
| FinWorkBench (Finch) | 73.3% | 2nd place among open models, trailing Kimi K3 |
| BrowseComp | 84.5% | Outperformed GLM-5.3 and DeepSeek Flash 4.1 1017 |
| B3 Agent Security Benchmark | 97.1% | Attack resilience (improved from 93.3% in the preview version) |
Terminal-Bench is a metric (Pass@1(%), higher numbers are better) that measures whether commands can actually be executed in a terminal to complete a given task to the end. It recorded 41% in Terminal-Bench 4.0, securing 1st place among open-weight models. Additionally, it scored 48% on the Artificial Analysis Coding Agent Index, achieving a score equivalent to the open-weight competitor model GLM-5.3 while successfully reducing the amount of tokens used per task by an average of 47%.
On the other hand, in human blind evaluations by Surge Code, while it outperformed other open-weight models such as GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash in terms of win counts, it did not reach the commercial model Claude Opus 5.5 among the evaluated targets. Furthermore, it recorded 73.3% on FinWorkBench (Finch) targeting financial workflows, but ranked 2nd among open models behind Kimi K3.
Note that the model supports three inference modes (Low, Medium, and the default High) so that developers can adjust the trade-off between performance and inference cost. By selecting according to the use case, more efficient operation is possible.
Strengths and Use Cases
Mistral Large 4 (Le Chonk) excels primarily in advanced agentic processing in the fields of software development (coding), defensive cybersecurity, and finance/knowledge work.
Coding and Autonomous Development Tasks
It possesses high token efficiency and accuracy in agent-type coding workflows, making it suitable for terminal operations and repository-wide code changes and maintenance. It is capable of autonomously advancing complex developments such as the following:
– Identification and fixing of cyber vulnerabilities: Reads reported vulnerability reports, reproduces bugs within an isolated sandbox, and repeatedly creates and adjusts fix patches until tests are passed.
– Autonomous web app development: Creates a 3D map app (Orbital Trace) that displays the orbits of approximately 19,000 artificial satellites and space debris in real time without human assistance, and even detects and fixes bugs and operational slowdowns in the browser by itself.
– Construction of scientific calculation programs: Can write a quantum chemistry solver from scratch that calculates electronic configurations within molecules, cross-checks results with existing standard tools, and builds a 3D interface to visualize molecular bond dissociation.
Defensive Cybersecurity
It has strengths in defensive security practices such as vulnerability discovery from codebases, operational verification, and repair without causing functional failures. While it does not cause unnecessary refusals for legitimate defensive analysis tasks, it maintains a high refusal rate for requests involving malicious cyber attack techniques. In addition, the following operation methods are provided for enterprises and public institutions:
– Enterprise-specific threat intelligence and detection logic can be trained using “Mistral Forge" to optimize the model for specific security tasks.
– Specialized variants specialized in penetration testing, vulnerability assessment, and red teaming are provided in preview to pre-vetted design partners (such as Orange Cyberdefense).
Financial and Practical Workflows and Secure Agent Operation
It supports spreadsheet creation and editing tasks that frequently occur in accounting and financial practices, as well as information gathering workflows using web browsing. In addition, because it has strong resilience against system prompt leakage and code injection attacks, AI agents deployed internally and externally can be operated safely. It is increasingly being utilized as a foundational model ensuring data sovereignty in European financial institutions (La Banque Postale, Crédit Agricole, ABN AMRO, Belfius, etc.) and space agencies (ESA Φ-lab).
How to Get It
Mistral Large 4 has its model weights released as open weights, allowing it to be deployed to custom infrastructure and self-hosted. This makes it possible to completely control data location and operating environments without depending on third-party APIs.
As an ecosystem to support the building and operation of this model, the following are provided:
– Mistral Forge: A toolset to customize and optimize the model using proprietary internal data, threat intelligence, and detection rules.
– Mistral Studio: A development environment to integrally build, govern, and improve operating rules and data access rights for created models and agents.
In addition to self-hosted operations, for organizations desiring managed cloud services, options such as usage through Microsoft Foundry and private deployment for Azure developers are also provided.
For details on the model release and introduction guidance, please check the Mistral Official Blog.
Related Articles
- Mistral AI Announces Mistral Large 4 with 1.05T Parameters
- Mistral AI Releases Timestral: Time Series Foundation Models
- Mistral AI Announces Voxtral 3 Audio Models and Voice Agent SDK
- Mistral Large 4 Released: Multimodal MoE Model
What to Read Next
- How to read Terminal-Bench → Benchmark glossary

