What is Le Chonk? Mistral Large 4 Explained
Le Chonk is the official nickname Mistral AI gave to Mistral Large 4 (ML4), a trillion-parameter multimodal language model announced on . It is Mistral's flagship model and, per the company, the largest AI model ever built in Europe.
The October 6, 2026 Announcement
On October 6, 2026, Mistral AI announced Mistral Large 4 as a public preview available through Mistral's API. The launch is staged: access is first rolling out to developers, cybersecurity professionals, and government authorities, with a broader public release scheduled for later in October 2026.
The announcement positioned the model around three pillars:
- Scale: 1.05 trillion total parameters, making it a trillion-parameter model for the first time at this scale from a European lab.
- Multimodality: native multimodal input through a 1.6-billion-parameter vision encoder, with a 1-million-token context window.
- Agentic operation: a hybrid system combining instruction following, reasoning, and agentic operation, positioned with a strong emphasis on cybersecurity.
On its intelligence leaderboard, independent tracker Artificial Analysis places the ML4 preview between DeepSeek V4.1 Flash and OpenAI's entry-level GPT-6 Luna.
Why Is It Called "Le Chonk"?
"Le Chonk" is not a community nickname that stuck after the fact — it is the name Mistral used in its own launch announcement. The informal title is a playful reference to the model's trillion-parameter size: this is a very large model, and "chonk" is internet slang for something big and heavy. The formal name remains Mistral Large 4, abbreviated ML4; "Le Chonk" is the affectionate shorthand Mistral itself introduced.
Mistral Large 4: Full Specifications
| Formal name | Mistral Large 4 |
|---|---|
| Nickname | Le Chonk |
| Abbreviation | ML4 |
| Total parameters | 1.05 trillion |
| Active parameters | 49 billion per token |
| Vision encoder | 1.6 billion parameters (native multimodal input) |
| Context window | 1 million tokens |
| Languages | Natively fluent in 160+ languages, including every official language of the EU |
| Architecture | Granular Mixture-of-Experts (MoE) |
| Operation modes | Hybrid: instruction following, reasoning, and agentic operation |
| API identifier | mistral-large-4 |
| Docs version | 26.10 (listed under the Open license category; docs page links to a playground) |
| Status | Public preview via API since October 6, 2026; open weights expected end of October 2026 |
The Training Story
Mistral trained Mistral Large 4 from scratch — no distillation. That distinction matters as context: Anthropic and OpenAI have accused Chinese competitors of distilling their models, a charge that has made "trained from scratch" a meaningful claim of provenance in the current AI race.
The training run took roughly two months on approximately 3,800 Nvidia Grace Blackwell GPUs — about 52 NVL72 racks — housed in Mistral's own European datacenters. The public preview is served from the same hardware.
The human story behind the model is one of scale-up. Mistral's science team grew from 3 researchers to roughly 300, and co-founder and chief scientist Guillaume Lample said the model's capabilities should improve further as reinforcement learning concludes and training capacity expands.
How Mistral Positions It
Mistral claims Mistral Large 4 is the best open-weight model from the US or Europe on aggregated benchmarks and competitive with the strongest open models worldwide. Cybersecurity is the centerpiece of that positioning: the model scores 93% on Cybench and 82% on CyberGym-E2E, and Mistral notes that several closed frontier models score near zero on those benchmarks because they refuse the task.
On the competitive front, CEO Arthur Mensch said the model is above Chinese models on certain aspects, including cyber. See the benchmarks page for the full numbers and the closed-frontier comparison for how it stacks up against proprietary models.
Current Availability
The API is live. Since October 6, 2026, Mistral Large 4 has been available as a public preview via Mistral's API (identifier mistral-large-4), with a staged rollout beginning with developers, cybersecurity professionals, and government authorities before a broader public release later in October.
The weights are not yet public. Open weights are planned for the end of October 2026 — Reuters reported October 27, 2026 — expected under a custom Mistral license and expected on Hugging Face. Until the weights ship, nobody outside Mistral can self-host the model.
- Pricing — preview API rates: $1.36 per 1M input tokens, $4.18 per 1M output tokens.
- Download — what to expect when the open weights ship.
- Quickstart — start calling the API today.