Mistral Releases Mistral Large 4 Open-Weight AI Model for Coding and Cybersecurity

What Happened

On 29 October 2026, Mistral released the weights for its new multimodal, multilingual Mixture-of-Experts (MoE) model, Mistral Large 4 (ML4), also known as “Le Chonk” [1, 2, 3]. The model features approximately 1 trillion total parameters with 49 billion active parameters and a 1-million-token context window [3]. It is available in two official variants: FP8 and NVFP4 [3].

Mistral aims to provide sovereign AI capabilities for enterprises and governments, allowing them to deploy the system on-premise within their own ecosystems [1, 2]. The model supports three reasoning modes to balance cost and performance: Low for simple tasks, Medium for everyday tasks, and High for complex, high-precision use cases [2].

In terms of performance, Mistral claims the model is comparable to the Chinese open-weight model GLM-5.3 on the Artificial Analysis Coding Agent Index, where ML4 scores 48% to 49.8% while using 47% fewer tokens per task [2, 3]. On the Artificial Analysis Cyber Index, ML4 scored 51% [2], while it achieved an 82% score on a specific test for reproducing and patching real vulnerabilities in open-source software [3]. Additionally, the model solved 93% of challenges in the Cybench security competition suite [3].

Access and pricing details include:

• The weights are available on Hugging Face and via the Mistral API [1].
• It can be used through Mistral Studio or OpenRouter [2].
• API pricing is set at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens [2].
• The model is distributed under the Mistral Ecosystem Open-Weight license, a modified MIT license [2]. This allows free use, modification, and commercialization, provided that businesses offering Model-as-a-Service must obtain a separate commercial license from Mistral [2].

Who Is Affected

• Businesses and Governments: These organizations can download and deploy the model to gain control over cyber defenses and avoid reliance on external AI providers [1]. Pierre Stock, Vice President of Science at Mistral, stated that open weights ensure access to intelligence cannot be cut off [1].
• Cybersecurity Professionals: The model is designed for defensive tasks such as vulnerability reproduction, patching, malware analysis, and writing detection rules [2, 3].
• Developers: The model provides tools for agentic coding and terminal-based workflows, where it reportedly improved by 13 percentage points on Terminal-Bench 4.0 [2].
• Researchers: Open weights allow researchers to directly measure refusal behavior rather than relying on provider descriptions [3].

Mistral’s CEO, Arthur Mensch, stated that AI should be a tool for empowerment rather than dominance [1]. Pierre Stock noted that providing these capabilities helps strengthen the ecosystem against malicious actors who might use similar technology to attack [1].