From Mistral · · 2m
Mistral Large 4 (Le Chonk) now open weight
Mistral releases open-weight 1-trillion-parameter model for coding and cybersecurity.
In brief
Mistral releases open-weight 1-trillion-parameter model for coding and cybersecurity. Mistral Large 4 (Le Chon: 1 trillion parameters) now open for developers with improved coding and cybersecurity capabilities. Originally reported by Mistral.
Mistral Large 4 (Le Chonk) released
Mistral released Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model now available with improved coding, cybersecurity, and reasoning efficiency.
Public preview success
The model entered public preview with strong results across coding benchmarks and has improved significantly in three weeks.
Coding benchmark leads
On Cognition's FrontierCode 2.0 benchmark, Mistral Large 4 leads with 19.7% score for code quality and mergeability.
Token efficiency gains
Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than leading open models.
Terminal benchmark improvement
On Terminal-Bench 4.0, Mistral Large 4 improved by 13 percentage points to 41% while reducing output tokens by 30%.
Three reasoning modes
Mistral Large 4 supports low, medium, and high reasoning modes for cost-performance balance in coding workflows.
Human preference wins
In Surge Code's blind evaluations, Mistral Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash.
Cybersecurity capabilities
Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index for vulnerability reproduction and patching without breaking functionality.
Refusal rates improve
Mistral Large 4 achieves 96.6% refusal rate on malicious cyber requests, the highest among open models.
Cybersecurity quotes
Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, and Rubrik all praise Mistral Large 4's cybersecurity capabilities.
Financial benchmark gains
Mistral Large 4 improved to 73.3% on Finch scores for finance-related workflows, second only to Kimi K3 among open models.
Web browsing performance
Mistral Large 4 reaches 84.5% on BrowseComp, starting from 75% in the SFT checkpoint.
Agent security benchmark
Mistral Large 4 resists 97.1% of attacks on the B3 Agent Security Benchmark, top of the OSS frontier.
Sovereign AI benefits
Mistral Large 4 gives organizations control to inspect, customize, and deploy on their own infrastructure with European training.
Enterprise adoption
La Banque Postale, ABN AMRO, Belfius, Crédit Agricole, ESA, Hugging Face, ZML, and Microsoft are adopting Mistral Large 4 for sovereign AI.
Deployment options
Mistral Large 4 is available through Mistral Studio, API, OpenRouter, Hugging Face, NVIDIA, and Microsoft Foundry.
The essence
Mistral Large 4 (Le Chon: 1 trillion parameters) now open for developers with improved coding and cybersecurity capabilities.
