Skip to content
Cyberttak.
· 2 min readBreaking

Mistral Large 4 opened its weights and scored 41% on Terminal-Bench 4.0 with 30% fewer output tokens

The 1-trillion-parameter model also scored 51% on a cyber index and refused 96.6% of malicious cyber requests, according to Mistral.

By CyberttakPublished · 15:04

In brief

Mistral opens a 1-trillion-parameter model

Mistral Large 4 now has open weights. It scores 41% on Terminal-Bench 4.0 with 30% fewer output tokens. Mistral says it refuses 96.6% of malicious cyber requests.

Two engineers work beside tall black server cabinets with open doors and cables along the floor in a bright data centre
Two engineers work beside tall black server cabinets with open doors and cables along the floor in a bright data centreCredit: AI-generated · FLUX.2 klein 9B

Mistral has released the weights of Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter model that handles text and images. The numbers below come from Mistral's own launch post, so they are the company's claims.

Weights are the learned settings of a model. Releasing them means companies can download the model, run it on their own machines and adapt it.

Coding

The model entered public preview earlier this month. Mistral says three more weeks of reinforcement learning, a training method that rewards good results, improved it.

  • On Cognition's private FrontierCode 2.0, it leads with 19.7%.
  • On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task.
  • On Terminal-Bench 4.0 it gained 13 points to reach 41%, with 30% fewer output tokens.

Tokens are the small chunks of text a model reads and writes, and fewer tokens means a lower bill.

In Surge Code's blind comparisons, people preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked ahead.

Reasoning modes

There are now three reasoning modes. They let developers trade cost against depth. One suits simple tasks and replaces the old 'None' mode. Another targets the hardest problems.

Security

On the Artificial Analysis Cyber Index the model scores 51%. That test covers finding vulnerabilities, reproducing them and patching them without breaking the code. Mistral says it avoids needless refusals on legitimate defensive work.

It also reports a 96.6% average refusal rate on malicious cyber requests, the highest it measures among open models. Booz Allen, Cisco, CrowdStrike, Cloudflare, ENISA and others commented, mostly praising open, self-hosted control.

Mistral also admits it saw potential for real attack uplift in live traffic, and is changing its guardrails. A specialised security variant goes only to vetted partners such as Orange Cyberdefense.

Limits: the benchmark results are Mistral's own reporting, and FrontierCode 2.0 is private. The excerpt cuts off before the finance results, and it gives no pricing.

Source: Mistral AI News