Full Breakdown
Mistral AI Unveils “Le Chonk,” a Trillion-Parameter Open-Weight Model
By Drooid · · How we work
Core Event: Launch of Mistral Large 4
On October 6, Mistral AI announced the preview of Mistral Large 4 (ML4), nicknamed “Le Chonk,” a one-trillion-parameter multimodal model optimized for coding, cybersecurity, finance, manufacturing and visual-grounding tasks. The preview is limited to developers, cybersecurity leaders and state authorities; full public weights are scheduled for release on October 27.
Background & Context: Open-Weight AI Competition
The AI field is split between closed models from U.S. firms and open-weight models, many of which are produced by Chinese labs. Europe has promoted a “third way” that emphasizes sovereign, openly available models. Founded in 2023 by former DeepMind and Meta researchers, Mistral has become Europe’s most heavily funded challenger, raising €3 billion in a Series D round at a €21 billion valuation.
Data & Statistics
- Model size: 1 trillion total parameters, 49 billion active parameters (sparse architecture).
- Training hardware: 4,000 Nvidia Grace Blackwell GPUs over roughly two months in European data centres.
- Language coverage: more than 160 languages, including every official EU language.
- Funding: €3 billion Series D (? $3.4 billion) completed last month.
- Customer base: supports over 125 global enterprises, including Airbus, ASML and HSBC.
Official Statements & Responses
- Lample also emphasized that “the cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”
Conflicting Reports & Gaps
Mistral’s materials cite a 62 % score on the DeepSWE v1.1 benchmark, positioning the model ahead of several Chinese rivals. Independent leaderboards, however, list higher scores for some of those rivals under alternative configurations, and the full benchmark methodology has not been publicly verified. The company has not disclosed which specific Chinese models or benchmark suites underpin the claim of cybersecurity superiority, leaving the comparison unsubstantiated beyond the CEO’s qualitative statement.
What’s Next
The preview period will run for roughly three weeks, during which selected partners will test the model with reduced safety restrictions. After the scheduled October 27 release, the model weights will be published under a custom Mistral license, enabling enterprises to run the model on sovereign infrastructure. Mistral plans to continue reinforcement learning and fine-tuning before the final checkpoint is made public, after which independent researchers can evaluate the model against established benchmarks.
