The short version
- Mistral Large 4 (Le Chonk) is an open-weight granular Mixture-of-Experts model with 1.05T total parameters and 52B active parameters, supporting native text-and-vision input, a 1M token context window, and 160+ languages.
- It ranks #1 among open-weight models on Cognition's FrontierCode agentic coding benchmark and sets a new Pareto efficiency benchmark on Artificial Analysis.
- Built for cyberdefense as well as coding, it was tested with security vendors including CrowdStrike, Cloudflare, Rubrik, and Zscaler, and leads open-weight models across cybersecurity, finance, and law.
- Weights are downloadable from huggingface.co/mistralai, and teams can customize the model on proprietary data with Mistral Forge for self-hosted deployment.
At a glance
| Architecture | Granular Mixture-of-Experts (1.05T total / 52B active parameters) |
|---|---|
| Context & Modality | 1M context window, native text + vision, 160+ languages |
| Coding Benchmark | #1 open-weight model on Cognition's FrontierCode |
| Cybersecurity Validation | Tested by CrowdStrike, Cloudflare, Rubrik, and Zscaler |
| Distribution & Tuning | Open weights on Hugging Face; customizable via Mistral Forge |
| Category | AI & Models |
| Creator / publisher | Mistral AI |
| Official link | huggingface.co ↗ |
| Heat | 🔥 797.1K Heat |
| Announced | |
| Data updated |
Inside the 1.05-trillion-parameter MoE that activates 52B per token
Mistral AI's flagship release pairs an official name — Mistral Large 4 — with the internal nickname that stuck, Le Chonk. Under the hood sits a granular Mixture-of-Experts architecture spanning roughly 1.05 trillion total parameters while routing each token through 52 billion active parameters at inference time. That split gives the model the knowledge capacity of a trillion-parameter frontier checkpoint while keeping per-token compute and latency closer to a mid-sized dense model.
According to Mistral's Hugging Face organization profile, Large 4 is trained from scratch as a natively multimodal generalist handling both text and vision across a 1-million-token context window and more than 160 languages. Artificial Analysis logged separate evaluations for Mistral Large 4 in Medium and High configurations on October 29, confirming that the 52B-active design shifts the Pareto frontier between intelligence and inference cost.
Top open rank on FrontierCode and field testing with cyber vendors
Mistral pitches Le Chonk around two high-stakes enterprise workloads: agentic software engineering and cyberdefense. On the coding side, the model ranks #1 among open-weight models on FrontierCode, the agentic coding benchmark maintained by Cognition. Rather than grading isolated single-function completions, agentic coding benchmarks stress multi-step repository navigation, tool invocation, and iterative patch verification.
On the defensive side, Mistral put the model through evaluations with four major cybersecurity and infrastructure companies — CrowdStrike, Cloudflare, Rubrik, and Zscaler — alongside domain benchmarks where it leads open-weight models in cybersecurity, finance, and law.
- Built to build (FrontierCode #1)
Leads all open-weight models on Cognition's FrontierCode benchmark while setting a new Pareto efficiency mark on Artificial Analysis.
- Built to defend (Cybersecurity)
Stress-tested on defensive security workflows by CrowdStrike, Cloudflare, Rubrik, and Zscaler.
- Built for sovereignty (Open weights + Forge)
Weights live on Hugging Face for local or private-cloud deployment, and Mistral Forge supports custom pre-training, post-training, and reinforcement learning.
Downloading the weights on Hugging Face and customizing with Mistral Forge
The model weights are published under the mistralai organization on Hugging Face at huggingface.co/mistralai, allowing enterprises and researchers to inspect the repository and deploy the checkpoint on their own hardware or preferred cloud provider. For regulated industries in finance, legal services, and defense that cannot send source code or incident logs to a third-party API, open weights turn a frontier-class model into an on-premises asset.
Teams that want to adapt Le Chonk beyond prompt engineering can pair the base weights with Mistral Forge. Mistral describes Forge as its platform for grounding custom AI in proprietary organizational knowledge across pre-training, post-training, and reinforcement learning, so security or engineering teams can specialize the 52B-active MoE on internal repositories and threat telemetry before deploying it behind their own firewall.
How to try it
Locate the weights on Hugging Face
Visit huggingface.co/mistralai and open the Mistral Large 4 (mistral-large-4) repository under the Frontier collection to review the model card and download the weights.
Provision inference for 52B active parameters
Set up your GPU cluster or inference endpoint with enough high-bandwidth memory to hold the 1.05T MoE weights while taking advantage of the 52B active parameter routing and 1M context window.
Customize with Mistral Forge if needed
Use Mistral Forge to run domain-specific post-training or reinforcement learning on internal codebases, legal corpora, or cybersecurity logs before production rollout.
FAQ
What is the relationship between Le Chonk and Mistral Large 4?
They are the same flagship release from Mistral AI: Le Chonk is the name used in Mistral's release announcement for the new open-weight model cataloged on Hugging Face as Mistral Large 4.
How many parameters does Mistral Large 4 (Le Chonk) have?
It uses a granular Mixture-of-Experts (MoE) architecture with roughly 1.05 trillion total parameters (listed as 1T total on the Hugging Face card) and 52 billion active parameters per token.
How does Le Chonk perform on coding and security benchmarks?
It ranks #1 among open-weight models on Cognition's FrontierCode benchmark, sets a new Pareto efficiency mark on Artificial Analysis, and was tested on cyberdefense tasks by CrowdStrike, Cloudflare, Rubrik, and Zscaler.
Can enterprises fine-tune and self-host the model?
Yes. The weights are available to download on Hugging Face for self-hosted deployment, and organizations can customize the model across pre-training, post-training, and reinforcement learning using Mistral Forge.
