Mistral has released the weights of an improved version of Large 4, its model nicknamed Le Chonk. The natively multimodal model has one trillion parameters and was previously available in public preview; developers can now download and run it on their own infrastructure.
Coding and reasoning modes
According to Mistral, another three weeks of reinforcement learning brought improvements particularly in coding and defensive security tasks. Developers can choose from three reasoning modes: Low replaces the previous None mode for simple tasks, Medium is intended for everyday work, and the default High mode is for the most demanding tasks requiring precision and thorough reasoning.
In the Terminal-Bench 4.0 coding test, Large 4 scored 41%, an improvement of 13 percentage points over the public preview, while producing 30% fewer output tokens, the company says. On the Artificial Analysis Coding Agent Index, it scored 48%, practically matching the open-weight GLM-5.3 while using an average of 47% fewer tokens per task.
Mistral reports a score of 19.7% on Cognition’s private FrontierCode 2.0 benchmark, which assesses overall code quality and whether a repository maintainer would accept an agent-generated change. In Surge Code’s blind pairwise evaluations, graders preferred Large 4 to GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash; it ranked behind Claude Opus 5.5 among the models evaluated.
Security tasks and a specialized variant
Large 4 scored 51% on the Artificial Analysis Cyber Index, according to Mistral. The index assesses finding vulnerabilities in code, reproducing and validating them, and patching them without breaking existing functionality.
Mistral demonstrated an agent that processes vulnerability reports, investigates source code and product context, and reproduces flaws in an isolated sandbox. It then revises a patch until a control exploit no longer works.
Selected enterprise partners have preview access to a specialized variant for penetration testing, vulnerability assessment and red-teaming. Access requires vetting and contractual safeguards, and testing is restricted to owned or explicitly authorized systems in controlled environments.
Results in finance and other tests
Mistral also reports improvements in creating and editing spreadsheets for finance and accounting. In FinWorkBench, known as Finch, the model scored 73.3% and placed second among open models behind Kimi K3, according to the company.
According to Mistral, Large 4 scored 84.5% on BrowseComp, compared with 75% for the starting checkpoint after supervised fine-tuning. Its result exceeded those of GLM-5.3 and DeepSeek Flash 4.1 1017. In the B3 Agent Security Benchmark, the model resisted 97.1% of attacks, compared with 93.3% for the preview version.
Download, deployment and license
The weights are available to download from Hugging Face, and developers can use vLLM or SGLang to run the model on their own infrastructure. Large 4 is also available through Mistral Studio, the Mistral API and Microsoft Foundry. Enterprises can tailor it to their own data and needs using Mistral Forge.
The Mistral Ecosystem Open-Weight license permits free use, modification, fine-tuning and commercial use. Businesses operating a Model-as-a-Service offering, however, need a separate license from Mistral for commercial use.
Pricing per million tokens is $0.70 for input, $0.07 for cached input and $2.10 for output.
