What Le Chonk actually is
On October 6, 2026, Paris-based Mistral AI launched the biggest model in its history: Mistral Large 4, nicknamed "Le Chonk." It is a Mixture-of-Experts model with roughly 1.05 trillion total parameters — but only about 49 billion are active for any given token. Think of it as a hospital with a thousand specialists on staff: each token is routed to just the few experts it needs. That design delivers frontier-scale reasoning at inference costs closer to a 50-billion-parameter model. It is natively multimodal, handling text and image inputs, with a dedicated 1.6-billion-parameter vision encoder for charts, documents, and satellite imagery.
Trained in Europe, on its own hardware
Mistral built the model from scratch on about 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers — not rented cloud capacity. The company says the training run took roughly two months and used two to three times less compute than Chinese rivals reportedly needed. The model supports more than 160 languages, including every official EU language, and offers a 1-million-token context window, though some preview providers cap it lower. For European policymakers, those details matter: this is frontier AI infrastructure built on European soil.
What it's actually for
Mistral is pitching Large 4 squarely at enterprises, with an unusual emphasis on cybersecurity — plus finance, legal, manufacturing, and visual grounding tasks like pinpointing objects in complex imagery. The model is live now as an API preview, listed at $1.36 per million input tokens and $4.18 per million output tokens. Independent evaluations back the cybersecurity and legal claims, put finance in the middle of the pack, and show a roughly fourfold jump over its predecessor on broad benchmark indexes. The company says the full open weights will be released by the end of October — the moment that turns a strong API product into a true open-weight milestone.
Why it matters: a third pole for open AI
For most of the past year, the strongest open-weight models have come from Chinese labs — DeepSeek, Qwen, Kimi, GLM — while the strongest models overall stayed closed behind US APIs. Mistral wants a third option: European, open-weight, and competitive. The launch is the first major milestone funded by Mistral's €3 billion Series D round, which valued the company at about €21 billion. But the open-weights promise is the real test. If the download lands as promised, researchers and companies worldwide will be able to run and fine-tune a trillion-parameter-class model on their own infrastructure.
How to try it today
You do not need to wait for the weights to experiment with Large 4. The API preview is live now through Mistral's own platform, and it has also been added to Vercel's AI Gateway, which puts it one API key away for many developers. Pricing starts at $1.36 per million input tokens and $4.18 per million output tokens, with cached inputs far cheaper. A few practical notes: the advertised 1-million-token context is not always available — some preview providers currently cap it around 524,000 tokens — and post-training work was still running at launch, so benchmark numbers may keep improving. When the open weights drop at the end of October, the same model becomes something you can download, fine-tune, and run on your own infrastructure.
If Europe's biggest open model ends up in your stack, what would you build with it?
Comments
Post a Comment