Mistral Large 4 Weights Released Today: 1-Trillion-Parameter Model Leads Open Coding Benchmarks

Mistral released the full model weights of Mistral Large 4 on November 9, 2026. The 1-trillion-parameter multimodal model achieved a score of 19.7% on the FrontierCode 2.0 benchmark.

What Happened

According to the official release, Mistral Large 4 had previously entered a public preview phase, and after three weeks of reinforcement learning its coding and defensive cybersecurity capabilities improved markedly. The model scores 19.7% on Cognition's FrontierCode 2.0 benchmark, up from 28% in the preview version to 41% on Terminal-Bench 4.0, and reaches 51% on the Artificial Analysis Cyber Index. It supports three reasoning modes—low, medium, and high—with the high mode as the default for high-precision scenarios.

How It Works

The model improves token efficiency through continuous reinforcement learning, scoring 48% on the Artificial Analysis Coding Agent Index while consuming 47% fewer tokens per task on average than comparable leading models. The three reasoning modes let developers choose based on task complexity, reducing inference costs. Trained from scratch on European infrastructure with a clear lineage, it provides a foundation for self-hosted deployments.

In blind testing on Surge Code, the model was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking second only to Claude Opus 5.5.

Industry Impact

Open weights combined with a modified MIT license make commercial use possible, though MaaS services require a separate license. This lowers the barrier for developers running complex coding agents while reducing unnecessary refusals in defensive cybersecurity tasks. Downloads are available on the Hugging Face platform, with support for vLLM and SGLang deployment.

Strategic Assessment (Analysis, Not Fact)

Based on the available benchmark data, Mistral Large 4 has established a local lead in open-model coding and cybersecurity, but whether its token-efficiency advantage translates into large-scale adoption still depends on actual deployment costs and ecosystem support. Its European training background may attract organizations focused on data sovereignty, but the overall gap with closed-source frontier models still requires continued observation.

• Three reasoning modes offer options balancing performance and cost
• Output pricing is $2.10 per million tokens
• Web-browsing task capabilities were not disclosed in detail in this release