Mistral Releases Mistral Large 4 for Self-Hosted Cyber Defense

Mistral AI has released Le Chonk, an AI model organizations can download, customize and run on their own systems for coding and cybersecurity work. Officially named Mistral Large 4, it has undergone three more weeks of training since its public preview.

The October 29 release comes with Mistral’s claim that it matches or surpasses leading Chinese open-weight models in coding and cybersecurity. The roughly one-trillion-parameter model accepts text and images and generates text and code. 

Coding Gains Depend on the Agent

Independent evaluator Artificial Analysis puts Mistral close to a current Chinese coding rival in its comparison of Mistral Vibe and Opencode. These are coding agents: software that lets a model inspect files, use tools and attempt programming tasks. The tested configurations pair Vibe with Large 4 at High reasoning effort and Opencode with GLM-5.3 at Max.

The two configurations scored 48 and 49 respectively on Coding Agent Index v2, which combines repository changes, terminal work and questions about codebases. Mistral’s configuration used fewer tokens and incurred a lower measured API bill per task attempt. Tokens are pieces of text, such as words, fragments or punctuation, that a model reads or generates.

Mistral attributes its improvement to further reinforcement learning, training the model on progressively harder tasks whose success can be checked. It reports that its Terminal-Bench 4.0 score rose from 28% in the preview to 41%, while output-token use fell 30%.

Broader performance still favors some rivals. On Artificial Analysis’ Intelligence Index v5, which covers a wider range of reasoning and knowledge tasks, Large 4 High scores 26 against 32 for DeepSeek V4.1 Flash 1017 Max. Moonshot AI’s Kimi K3.1 Pro scores 46, but its weights were not public on October 29: it offered hosted API access rather than the same self-hosting option.

What the Cyber Tests Measure

The defensive workflow starts with software source code. An agent examines it for flaws, investigates a possible failure and proposes a repair. In Mistral’s vulnerability-triage demonstration, the agent receives a vulnerability report, investigates the relevant code, reproduces the failure in an isolated environment and iterates on a patch until a test attack stops working.

Artificial Analysis’ Cyber Index combines three equally weighted evaluations: code audits and patching, scanner-guided vulnerability discovery, and crash reproduction followed by repair. Each begins with source access.

Mistral reports 51 points on that index. DeepSeek V4.1 Flash at 57 and Xiaomi’s MiMo-V2.6-Pro rank at 56.

The crash-and-patch component, CyberGym-E2E-AA, uses 131 tasks from 131 C/C++ projects, with 90 minutes per attempt. A pass requires an input that crashes the original program, a patch that stops that crash, and passing functionality tests. An agent can therefore earn a pass by repairing a different real crash from the one that prompted the task.

Control Comes With Responsibility

Pierre Stock, Mistral’s vice president of science, highlighted to WIRED how thir downloadable weights give enterprises and governments continued access even if a provider stops offering a service. He also raised the possibility of Chinese models behaving maliciously, framing a Western alternative as a security concern.

Rafe Pilling, senior director of threat intelligence at security company Sophos, told WIRED that the same freedom to download and customize a model enables both productive use and abuse. He said an internet-connected model could siphon data, but advanced models from any country need careful supervision. Monitoring and controls can help operators detect deceptive behavior.

Mistral itself acknowledges potential for the model to assist real attacks and says it will strengthen guardrails over the coming months. A separate variant for advanced security work, including penetration testing, remains in preview for vetted, contracted partners. That testing probes systems through authorized attacks; partners are restricted to systems they own or have explicit permission to test, in controlled environments.

Reasoning Effort and Commercial Terms

The release adds Low and Medium reasoning effort alongside High, the default, letting users reduce the work the model spends on an answer. That changes capability as well as cost: Artificial Analysis’ Vibe results put Medium at 35 index points and $0.78 per task attempt, against High’s 48 points and $1.26.

Mistral’s prompt cache reuses computation when a new request begins with previously processed input. Its API rates are $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. The benchmark task bills reflect the mix and quantity of those tokens.

Commercial permissions also depend on the license. Mistral Large 3 shipped under Apache 2.0 in December 2025. Large 4 uses the Mistral Ecosystem Open-Weight License, a modified MIT license. If a licensee or an affiliate operates a Model-as-a-Service business, the licensee needs a separate commercial agreement before using Large 4 commercially.

The license defines that business as giving third parties access to run the model or train it further, with meaningful control over inputs, parameters or training data. Products that only embed model capabilities in specific features, and services that merely relay requests to models hosted elsewhere, are excluded from that definition. Internal use and access through Mistral’s official products are exempt from the separate-agreement requirement.