Mistral releases the weights of its flagship Mistral Large 4, aka "Le Chonk"

Mistral AI has released the weights of Mistral Large 4. The company's flagship natively multimodal model has one trillion parameters and goes by the affectionate nickname Le Chonk. It entered public preview earlier this month. Since then, Mistral has applied user feedback and kept scaling reinforcement learning. The company says gains are still accelerating, with no sign of saturation.

Under the hood it is a hybrid MoE model that combines instruction following, reasoning and agentic capabilities in one system. It accepts multimodal input and, according to Mistral, is natively fluent in more than 160 languages.

Coding

The model received three more weeks of reinforcement learning after the preview. It now leads Cognition's private FrontierCode 2.0 benchmark with 19.7%; that test checks whether a repository maintainer would merge the agent's change. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the top open-weight model, GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it gained 13 percentage points over the preview, reaching 41% and taking the lead among open-weight models. Output tokens dropped by 30% along the way.

To balance quality against cost, the model offers three reasoning modes: Low for simple tasks, Medium for everyday work and High for the hardest problems. High is the default. In Surge Code's blind pairwise evaluations, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher.

Cybersecurity

The model scores 51% on the Artificial Analysis Cyber Index. The index measures how well a model finds vulnerabilities, reproduces them and patches them without breaking anything. Built-in guardrails are meant to separate defensive tasks from offensive ones. The model refuses malicious requests 96.6% of the time on average, which Mistral says is the highest rate it has measured among open models.

Mistral acknowledges that it monitored live traffic after the preview. The model refused black-box attacks against live targets, but the company still sees potential uplift for real attacks. It will keep adjusting its guardrails over the coming months. A specialized variant for penetration testing and red-teaming is going only to vetted partners such as Orange Cyberdefense, under contractual safeguards. The announcement includes endorsements from Booz Allen, Cisco, Cloudflare, CrowdStrike, ENISA and others.

Other workloads

On the Finch finance benchmark the model scores 73.3%, second only to Kimi K3 among open models. On BrowseComp it rose to 84.5%, up from 75% for the SFT checkpoint. On the B3 agent security benchmark it resisted 97.1% of attacks, compared with 93.3% in the preview.

License and availability

The model was trained from scratch on European infrastructure. Mistral positions it as a foundation for "sovereign AI": organizations can deploy and fine-tune it on their own infrastructure without relying on Mistral's API. The weights are on Hugging Face, with vLLM and SGLang suggested for serving. The model is also available in Mistral Studio, through the Mistral API, on OpenRouter, in Microsoft Foundry and as an NVIDIA NIM container.

The license, called the Mistral Ecosystem Open-Weight license, is a modified MIT license. It allows free use, modification and commercialization. A separate commercial license is required only for businesses offering the model as a service (Model-as-a-Service). API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Source: Hacker News