Mistral Large 4 (Le Chonk): Specs, Benchmarks and Open Weights

What Is Mistral Large 4?

Mistral's launch post calls Mistral Large 4 the company's largest and most capable model to date: an open-weight, natively multimodal mixture-of-experts model.

Mistral's own sources describe its size two ways, so the table below gives both, with the Mistral Large 4 model card as the source for the technical details:

Mistral says it will share more on the architecture and its post-training method before the weights ship, so treat the preview specs as subject to change.

How Does Mistral Large 4 Perform on Benchmarks?

Mistral claims the model is the strongest open-weight model built in the US or Europe on aggregated benchmarks and state of the art among open models on cybersecurity, finance and legal work. These are the scores Mistral published for the preview:

Read these as vendor numbers on a preview model. In its coverage of the launch, CNBC notes that the model still lags the frontier in areas such as coding.

The LLM benchmarks guide covers what each kind of benchmark can and cannot tell you about your own workload.

Why Is Mistral Large 4 Pitched for Cyber Defense?

Mistral's argument is about what the model will do as much as how well it scores. The launch post says closed models including Claude Opus 5.5 and GPT-6 Astra score near zero on that reproduce-and-patch test because they refuse the task, even though defending software often starts with proving a flaw is real. Mistral lists what the model handled in its own testing:

• Security operations work - analysing malware, prioritising vulnerabilities and writing detection rules, beyond what it was explicitly trained for.
• Self-deployment - running on private cloud or on-premise, so security teams can use it under their own policies without a provider-level refusal mid-incident.
• Pre-release red-teaming - until the weights ship, cybersecurity leaders, vetted partners and state authorities are testing the same model with reduced moderation and expanded cyber capabilities.

Co-founder and chief scientist Guillaume Lample told CNBC that the cyber defense capabilities "will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks." CNBC links the demand to the OpenAI Hugging Face incident, where Hugging Face's defense was blocked by closed models' guardrails and it turned to an open model instead.

When Will Mistral Large 4's Open Weights Be Released?

Mistral says the weights drop at the end of October 2026, and its announcement on X repeats the date. Until then the model is a public preview API, served from Mistral's European infrastructure. Open weights change who controls the model:

• Hosting - you can run the model in your own cloud or datacenter instead of calling a provider's API.
• Modification - CNBC notes that open models can be modified and self-hosted, in contrast to the leading closed systems from OpenAI and Anthropic.
• Sovereignty - Mistral also runs a European deployment end to end, independently of other digital service providers and under European law.
• Guardrails - when you host the weights, moderation is whatever you put around the model, so jailbreak and prompt-injection defenses become your team's job to build and test.

Mistral and TestMu AI are both signatories of the Open Weights and American AI Leadership letter; the open weights letter post explains why engineering teams need models they can inspect, run and adapt.

How Do You Test an App Built on Mistral Large 4?

Moving an agent onto a new model is a release, even when no line of your code changes: the same prompt can produce different answers, refusals and tool calls. With TestMu AI, Agent Testing grades what a chat or voice agent says and Agent Assurance checks what a tool-using agent does, so each risk below has a matching check:

Agent Testing: Grade What the Agent Says

Agent Testing uses AI testing agents to hold multi-turn conversations with your chat, voice or phone agent and score every reply, before and after deployment.

• Regression after a model update - scenarios carry a version history, so you can rerun the same versions when the underlying model changes and see drift as a score change instead of a customer complaint.
• Standard metrics - 9 metrics for chat and voice, including hallucination, bias, completeness and context awareness.
• AI red teaming - AI red teaming runs prompt injection, jailbreak and data exfiltration attacks across 9 attack categories at 3 intensity levels and grades each category from A+ to F.
• Multilingual scenarios - scenario generation, execution and evaluation are multilingual, which matters for a model trained on more than 160 languages.

Run the same suites from your terminal and CI with the Agent Testing CLI, and see agent regression testing for how to structure a before-and-after comparison.

npm install -g @testmuai/rook

Getting Started With Mistral Large 4

Try the preview API on Mistral Studio with the model ID mistral-large-4, and estimate prompt sizes for your workload with TestMu AI's free token counter before you plan around the 1M-token window.

Before you switch a production agent, run your regression and red-team suites against both the current model and Mistral Large 4, and compare the results. If you plan to self-host when the weights ship, schedule the same runs against your own deployment, because your guardrails are now part of what you are testing.