Mistral Opens Weights of Its Trillion-Parameter Model: Coding Agents Use 47% Fewer Tokens, Terminal Tasks Gain 13 Points
Synopsis
Mistral AI has released the weights of its flagship Mistral Large 4 (Le Chonk), a trillion-parameter natively multimodal model improved by three more weeks of reinforcement learning after its public preview. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average; on Terminal-Bench 4.0 it gains 13 percentage points over its own preview to 41% with 30% fewer output tokens. Three reasoning modes let developers trade performance against inference cost.
Interpretation
Mistral AI has released the weights of its flagship Mistral Large 4, letting organizations inspect, customize and deploy the trillion-parameter model on their own infrastructure instead of depending on a vendor API. The model had been available only as a public preview without open weights; the release ships an improved version after three additional weeks of reinforcement learning. On Cognition's private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge the agent's change, Mistral Large 4 leads with a score of 19.7%.
The coding gain shows up in token consumption: at a level effectively tied with the leading open-weight model, the model uses 47% fewer tokens per task on average, making coding agents cheaper to run at scale. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, effectively tied with GLM-5.3, with the difference lying in token use. On Terminal-Bench 4.0, the model improved by 13 percentage points to 41% over its public preview while cutting output tokens by 30%, taking the lead among open-weight models.
The model adds low, medium and high reasoning modes so developers can trade performance against inference cost by task difficulty, with high as the default. Low replaces the previous none reasoning mode for simple tasks, medium targets everyday tasks, and high targets the hardest cases needing high precision and thorough reasoning. In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred over open-weight models including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated.
Perspective
The report covers Mistral AI's release information and vendor-reported evaluations: benchmark results across coding, terminal workflows, cybersecurity, finance spreadsheets and web browsing, plus the three reasoning modes and open-weight license terms. It is aimed at developers working on open-weight deployment, coding-agent cost and defensive security tooling.
Mistral AI says it observed the model refusing black-box attacks against live targets in live traffic, yet sees potential for uplift to conduct real attacks, so it is modifying its guardrails and will keep updating them in the months to come. The specialized cybersecurity variant is available only to rigorously vetted enterprise design partners, with testing restricted to owned or explicitly authorized systems in controlled environments.