AI Tech Flash ── 2026-10-30
Mistral Releases Le Chonk Weights, Claims Top-Tier Open Model
Mistral releases the trillion-parameter Le Chonk; it claims top-tier coding results.
What changes
According to Mistral, the weights of Mistral Large 4 (ML4), nicknamed Le Chonk, are now public. Weights are the numerical data a model learns in training. Anyone can download them and run the model on their own systems.
According to ZDNET, the model's coding scores are close to those of leading Chinese open models. That suggests Western open models are narrowing the gap with Chinese ones.
Why it matters
According to WIRED, a series of AI-related cybersecurity incidents has occurred. Mistral argues that the West therefore needs its own advanced open-weight models.
Mistral's Pierre Stock said such models differ from ones whose provider can cut off access remotely. Rafe Pilling of the security company Sophos said open models are dual-use technology that can also be abused.
What it is
Le Chonk is a large language model from Mistral, formally named Mistral Large 4 (ML4). According to ZDNET, the nickname refers to its trillion-parameter size. Parameters are the adjustable numbers that determine a model's behavior.
According to WIRED, Mistral calls the finished version the "world's best" open-weight model for automated coding and cybersecurity. According to ZDNET, it offers three reasoning modes: low, medium, and high.
Who and where
The French AI company Mistral built the model. According to ZDNET, the preview shipped on Oct. 6, and the weights have now been released.
According to WIRED, the final version is freely available to anyone. According to ZDNET, it was trained in Mistral's own data centers in Europe.
Why it was built
Mistral told WIRED that the West needs its own advanced open-weight models. Stock said there is no guarantee that Chinese models will not behave maliciously.
According to ZDNET, Mistral says the model should help enterprises and governments defend against attackers who jailbreak closed models to run cyberattacks. Pilling of Sophos said that advanced models from any country can go rogue without careful supervision.
How it works
According to ZDNET, ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months. GPUs are chips built for parallel computation. A select group of partners then tested it for three weeks after the preview.
According to WIRED, Mistral says it fed the model increasingly difficult cybersecurity and coding problems. According to ZDNET, testing partners received a version with fewer guardrails and expanded cybersecurity capabilities.
Key points
- According to ZDNET, the Terminal-Bench 4.0 score rose from 28% in the preview to 41%, one point above GLM-5.3.
- According to ZDNET, the overall index scores are 48 for ML4, 49 for GLM-5.3, and 48 for Codex GPT-6.1 Terra.
- In Mistral's statement, it rivals GLM-5.3 while using 47% fewer tokens per task on average.
Conclusion
Mistral claims Le Chonk is the best open-weight model for automated coding and cybersecurity. In ZDNET's reporting, it trails GLM-5.3 by one point on the overall index. ZDNET cautions that benchmark scores should be taken with a grain of salt.
Quotes and sources
“It’s a pivotal moment in the global AI leadership race,”── WIRED, 2026-10-29
“It’s clearly a dual-use technology,”── WIRED, 2026-10-29
“ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral’s own data centers in Europe,”── ZDNET, 2026-10-30
“rivals top open models such as GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% less tokens per task on average”── ZDNET, 2026-10-30
Source articles
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.