Skip to content
MaungaAI news, explained

Models

Mistral lets anyone download its biggest AI and sells it as a shield against cyberattacks

French company Mistral has released Le Chonk so that businesses and governments can install it on their own machines, where no one can cut off their access. The catch: attackers can download it too.

2 min read5 sources, one officialWritten with AI · machine-translated

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • Since October 29, anyone can download Mistral Large 4, nicknamed Le Chonk, and run it on their own computers (Mistral AI News, TechCrunch, Hugging Face Blog).
  • It has one trillion parameters, the settings the AI learns, and can read one million tokens, or pieces of words, in one go (Hugging Face Blog, Mistral AI News).
  • Claimed by one partyAccording to Mistral, it codes as well as the Chinese open model GLM-5.3 but uses 47% fewer tokens per task (Mistral AI News, TechCrunch).
  • Claimed by one partyMistral says it refuses 96.6% of malicious cyberattack requests, the highest figure among the open models it has measured (Mistral AI News).

Why it matters

In open models, the benchmark until now has been Chinese labs such as DeepSeek and Z.ai (WIRED). If Mistral's figures hold up, businesses and governments would have a European option for finding and fixing security flaws on their own machines, without relying on another company's service (Mistral AI News, WIRED). The price is that it is also within reach of attackers (WIRED).

The details

Mistral showed a preview version in early October and released the final one on October 29, with improvements in coding and cybersecurity (Mistral AI News). It comes with three reasoning modes: more reasoning gets more answers right, and less reasoning costs less (Mistral AI News). The license is based on MIT, one of the most permissive, but it sets limits for anyone who wants to offer the model as a service to others (Hugging Face Blog).

Pierre Stock, Mistral's vice president of science, argues in WIRED that an open model cannot be switched off remotely, unlike those from OpenAI, Anthropic or Google (WIRED). Rafe Pilling, of security company Sophos, points out in the same outlet that it can be used both to attack and to defend (WIRED). There is also a mismatch: Hugging Face and Mistral publish different results for the model on the same cybersecurity test (Hugging Face Blog, Mistral AI News).

What we don't know

  • Whether anyone outside Mistral confirms that it performs as the company says.
  • What its real score is on the cybersecurity test, since Mistral and Hugging Face give different figures.
  • How much it costs to use through Mistral's paid service.

Sources

This story is based on 5 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original

The thread: Mistral · Mistral Large 4

  1. Mistral lets anyone download its biggest AI and sells it as a shield against cyberattacks · you are here
  2. Mistral's new open model can't be downloaded yet