Mistral AI Releases Open Weights for 1-Trillion Parameter Mistral Large 4

The Three-Question Framework

What Happened?

On October 29, 2026, Mistral AI officially released the open weights for Mistral Large 4 (internally codenamed 'Le Chonk'), its flagship 1-trillion parameter natively multimodal foundation model. Following three weeks of intensive reinforcement learning scaling since its initial preview, the weights are now available for download on Hugging Face under the Mistral Ecosystem Open-Weight license. The model introduces three explicit runtime reasoning modes (Low, Medium, and High) and takes the lead among open-weight architectures on Cognition's FrontierCode 2.0 benchmark with a 19.7% score.

Why It Matters

The release establishes an unprecedented milestone for open-weight artificial intelligence, delivering frontier agentic coding and defensive security capabilities that previously required proprietary hosted APIs. On the Artificial Analysis Coding Agent Index, Mistral Large 4 ties leading proprietary models at 48% while consuming 47% fewer tokens per task. Furthermore, on the Artificial Analysis Cyber Index, it achieves a score of 51% in autonomous vulnerability reproduction and patch synthesis while recording a 96.6% refusal rate against malicious offensive requests.

What Could Happen Next

Enterprises can self-host the open weights immediately using vLLM or SGLang on local clusters, deploy optimized NVIDIA NIM containers across TensorRT-LLM runtimes, or access hosted endpoints via Microsoft Foundry, OpenRouter, and the Mistral API. API pricing has been established at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Mistral has confirmed that a comprehensive technical architecture report will follow shortly.

Executive Key Takeaways

Mistral AI releases open weights for 1T-parameter multimodal Mistral Large 4 under its Ecosystem Open-Weight license.

Takes #1 open-weight ranking on Cognition FrontierCode 2.0 (19.7%) and Terminal-Bench 4.0 (41%), while matching 48% on Coding Agent Index using 47% fewer tokens.

Scores 51% on Artificial Analysis Cyber Index for vulnerability reproduction and patching with a 96.6% refusal rate on malicious requests.

Source: Mistral AI official announcement published October 29, 2026, titled 'Mistral Large 4 (Le Chonk) is now open weight'.

Trained from scratch on European sovereign compute infrastructure, Mistral Large 4 unifies instruction following, deep chain-of-thought reasoning, and agentic tool invocation across text and vision. By scaling reinforcement learning algorithms beyond initial preview baselines, engineers achieved substantial efficiency breakthroughs in code synthesis and defensive cybersecurity without observing performance saturation. In blind side-by-side evaluations conducted by Surge Code, human evaluators preferred Mistral Large 4's outputs over GLM-5.3 (64% to 36%), Kimi K3 (64% to 36%), and DeepSeek V4.1 Flash (56% to 44%).

Defensive cybersecurity capabilities formed a core pillar of the post-training curriculum. In enterprise evaluations conducted with cybersecurity partners including Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, and Orange Cyberdefense, the model demonstrated high precision in triaging vulnerability advisories, isolating code flaws in sandboxes, and generating verified patches. On the B3 Agent Security Benchmark, Mistral Large 4 resisted 97.1% of adversarial prompt injections, up from 93.3% during early preview rounds.

The model is licensed under the Mistral Ecosystem Open-Weight license, an MIT-derived structure granting royalty-free usage, modification, and commercialization, with the sole exception that commercial Model-as-a-Service providers must secure a dedicated agreement. In addition to downloadable weights, managed serving is live across Microsoft Azure Foundry and NVIDIA cloud endpoints.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations