Mistral has released the weights for Mistral Large 4 (ML4) on Hugging Face, describing it as its flagship open frontier system, per IT Brief Asia.

The model is a natively multimodal, 1-trillion-parameter system with 52 billion active parameters, available both as a download and through Mistral's own API.

Mistral president of science Pierre Stock said two reasoning modes were added during the same RL run, with the model offering low, medium and high reasoning levels, and no signs of saturation.

The positioning is unusual: rather than a general chat model, Mistral is pitching ML4 at coding and cybersecurity workloads for organizations that want to run models on their own systems, IT Brief Asia reports.

That appeals to regulated sectors seeking tighter control over code, data and access, and the model can be deployed on customer infrastructure for cyber defence work so enterprises and governments can detect and respond to threats without depending on a third party.

Mistral also outlined industrial uses, from turning 2D blueprints into 3D CAD models for manufacturers to spreadsheet and filing-extraction work for financial institutions, plus geospatial analysis of satellite imagery for wildfire and infrastructure monitoring.

The evidence rests partly on vendor-cited figures.

According to IT Brief Asia, Mistral said ML4 ranks as the top open model for agentic coding on Artificial Analysis' Coding Agent Index, lifted its Terminal-Bench 4.0 score from 28% to 41% to level with GLM-5.3, and uses 47% fewer tokens per task than GLM-5.3 at similar quality.

The broader field is crowded: shattered.io's spec sheet places Kimi K3, DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3 and Mistral Large 4 side by side, noting licences differ sharply — DeepSeek ships under MIT and Qwen3.8's 27B under Apache 2.0, while Mistral's modified MIT licence carries a Model-as-a-Service restriction.

shattered.io also lists ML4 at roughly $0.70 input and $2.10 output per million tokens with a $0.07 cached-input rate.

The release lands in an intensifying debate over whether leading systems should ship weights or stay behind APIs, and Mistral framed it as an efficiency argument: CEO Arthur Mensch said other labs raised hundreds of billions while Mistral raised roughly $6 billion in three years, claiming global leadership in compute efficiency.

shattered.io reports the model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centres, per IT Brief Asia's account of the company's claims.

For buyers, the takeaway is that deployment choice now matters as much as leaderboard rank — shattering.io's analysis stresses licence rights, data residency and cached-token pricing as the real cost drivers once a pilot becomes a product, with Kimi K3.1 Pro weights reportedly due November 4, keeping the field in motion.