Mistral has released a public preview of its new trillion‑parameter model, Mistral Large 4 (ML4) – nicknamed “Le Chonk”. The model, trained on 4,000 Nvidia Grace Blackwell GPUs over two months in Europe, is positioned as an open‑weight solution that lets enterprises keep full control of their cyber‑defense capabilities. Mistral plans to publish the model weights on 27 October, after a month of testing with developers, security leaders and state authorities.
Key Takeaways
- ML4 (“Le Chonk”) is a trillion‑parameter open‑weight model focused on cybersecurity.
- It was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs in two months, a fraction of the compute used for OpenAI’s GPT‑6 Astra.
- Mistral emphasizes data sovereignty, offering self‑deployment or API access in a European‑jurisdiction region.
- Early third‑party tests show ML4 matching or exceeding proprietary models on computer‑vision and financial tasks, and reaching 15 % on the Harvey Legal Agent benchmark for open‑weight models.
- Weights will be released on 27 October after a public‑preview period.
| Metric | Value | Why it matters |
|---|---|---|
| Model name | Mistral Large 4 (ML4) – “Le Chonk” | Identifies the new open‑weight AI offering. |
| Parameter count | ~1 trillion | Places the model in the same scale as leading LLMs. |
| Training compute | 4,000 Nvidia Grace Blackwell GPUs for two months | Demonstrates efficiency compared with rivals that use ~100,000 GPUs. |
| Weights release date | 27 October 2024 | Marks when developers can fully own the model. |
What is Le Chonk?
Le Chonk is Mistral’s latest open‑weight large language model, built from the ground up to serve cybersecurity, finance and multimodal workloads. By publishing the weights, Mistral lets customers run the model on‑premises or via a regional API, preserving data sovereignty and avoiding vendor lock‑in. The model’s trillion‑parameter size puts it in the same class as other “large” LLMs, but its open nature differentiates it from closed offerings such as Anthropic’s or OpenAI’s proprietary suites.
Security‑first design philosophy
Mistral frames ML4 as a “democratic defender” against AI‑driven attacks. In a briefing, co‑founder Guillaume Lample said:
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks.”
The company argues that open models can be audited, customized, and, crucially, cannot be arbitrarily disabled by a vendor—a risk it associates with closed‑source services. Mistral also joined Nvidia’s Open Secure AI Alliance, a coalition that promotes open‑weight tools as a means to democratize AI security.
Performance and early benchmarks
Third‑party analysis, though still pending final numbers, suggests ML4 competes with higher‑priced proprietary models. In computer‑vision tasks, it performed on par with OpenAI’s GPT‑6 Astra. On financial workloads, it matched or slightly outperformed DeepSeek models. Notably, it achieved a new high of 15 % on the Harvey Legal Agent benchmark for open‑weight models, indicating strong reasoning abilities in legal contexts.
Mistral’s spokesperson told ZDNET that “in absolute terms on cyber capabilities, ML4 outperforms the best models from Kimi, Deepseek and Meta.” The claim underscores the model’s focus on threat detection, malware analysis, and other security‑centric tasks.
Compute efficiency versus the competition
Training a trillion‑parameter model with 4,000 GPUs over two months is a stark contrast to the roughly 100,000 GPUs Nvidia’s CEO Jensen Huang said OpenAI used for GPT‑6 Astra. This efficiency reflects Mistral’s smaller scale and European data‑center footprint, but the company expects rapid improvements as it iterates on the base model.
Deployment options and data sovereignty
Mistral promises flexible deployment: customers can self‑host the model in their own infrastructure or consume it through an API hosted in a region of their choice. A dedicated “European sovereign region” will keep data under EU jurisdiction, addressing regulatory concerns that have become prominent after incidents like the Hugging Face breach.
Why it matters
Le Chonk offers a rare combination of scale, open‑weight accessibility, and a security‑centric design. For enterprises and governments wary of relying on closed AI services that could be withdrawn or manipulated, ML4 provides a path to own the model, audit its behavior, and tailor defenses to specific threat landscapes. Its relatively modest compute footprint also signals that high‑performance AI can be built without the massive resource pools of the biggest U.S. labs, potentially reshaping the economics of AI security.
Frequently asked questions
Who developed Le Chonk?
Mistral, a French AI laboratory, created the model. Co‑founder Guillaume Lample and VP of Science Pierre Stock led the public briefing.
When will the model weights be publicly available?
The weights are scheduled for release on 27 October 2024, after a month‑long public preview.
What makes ML4 different from proprietary models?
ML4 is open‑weight, allowing full user control, self‑deployment, and avoidance of vendor lock‑in. It also emphasizes data sovereignty with a European‑jurisdiction deployment option.
How does the training compute compare to rivals?
ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs for two months, whereas OpenAI’s GPT‑6 Astra reportedly used about 100,000 GPUs.
Can Le Chonk be used for non‑security tasks?
Yes. Mistral highlights strong performance in finance and multimodal use cases, though final benchmark numbers are still pending.
Bottom Line
Mistral’s Le Chonk delivers a trillion‑parameter, open‑weight AI model that prioritizes cybersecurity while offering the flexibility to run the model on‑premises or via a regional API. Its efficient training footprint and early benchmark results position it as a viable alternative to costly proprietary solutions, especially for organizations that demand data sovereignty and full control over their AI defenses.
Reporting based on: original source.
Comments
Post a Comment