Mistral Releases Open Weights for Trillion-Parameter 'ML4' Targeting Cybersecurity and Coding Agents
French AI firm Mistral has officially released the weights for Mistral Large 4 (ML4), its trillion-parameter foundation model. Available via Hugging Face and AP
French AI company Mistral officially distributed the model weights for Mistral Large 4 (ML4), its trillion-parameter flagship model, on November 12, 2026. Initially previewed earlier this month under the codename "Le Chonk," the model is distributed under an open-weight release that allows enterprises and public sector institutions to host and run the architecture directly on local infrastructure without external cloud dependencies.

Image source: https://www.kucoin.com/news/flash/mistral-releases-trillion-parameter-ai-model-for-cybersecurity
The weights are accessible through the Hugging Face repository and Mistral's native API. Aimed directly at enterprise security operations and government agencies that restrict sending sensitive data to closed-source providers such as OpenAI and Anthropic, ML4 centers on sovereign cyber defense capabilities where organizations retain full ownership of their operational telemetry and model execution.
47% Fewer Tokens than GLM-5.3 and Top Ranking on Coding Agent Index
Mistral highlighted inference efficiency as a core competitive differentiator for ML4. Compared against China's GLM-5.3—previously recognized as a primary open-weight benchmark for cybersecurity workloads—ML4 requires 47% fewer tokens per task. This reduction significantly lowers total operational costs for organizations running persistent threat monitoring pipelines.
The model also marked strong results in automated software engineering. On Artificial Analysis's Coding Agent Index benchmark, ML4 outperformed all other open-weight models on autonomous agent programming tasks. Reaffirming the company's release philosophy, Mistral CEO Arthur Mensch stated that AI should "be an empowering tool, not a dominating one," arguing that sovereign, open architectures can deliver industry-leading performance without the capital-intensive compute footprints common to closed-source labs.
Sovereign Cyber Defense and Governance Across Autonomous Agents
Releasing raw model weights addresses critical autonomy requirements in defensive infrastructure. Pierre Stock, Vice President of Science at Mistral, told TechCrunch that because the weights are open, access to this intelligence cannot be revoked or throttled by external cloud operators. This setup allows defense teams to deploy ML4 entirely within isolated environments to identify, coordinate, and respond to major security incidents in real time.
However, distributing high-parameter models publicly introduces clear security tradeoffs. Advanced capabilities that assist defenders in locating software vulnerabilities at rapid speed can also be adopted by threat actors operating in closed networks. Addressing broader industry concerns regarding rogue or out-of-control AI agents, Stock acknowledged the operational challenges while cautioning against panic. He maintained that strengthening collective defense across the wider software ecosystem as quickly as possible remains the most effective hedge against emerging automated threats.
Sources
- KuCoin News Flash (citing CoinDesk and TechCrunch): Mistral Releases Trillion-Parameter AI Model for Cybersecurity