Mistral's 'Le Chonk' lands: a 1-trillion-parameter multimodal model with open weights promised within the month

What Mistral Large 4 actually is

On October 6, 2026, Mistral launched a public preview of Mistral Large 4, its largest and most capable model to date — internally nicknamed le Chonk. It is a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters, of which only about 49 to 52 billion actively fire per token. It accepts both text and image input, runs a 1-million-token context window, supports 160+ languages including every official EU language, and is served in Mistral Studio under the model ID mistral-large-4.

One caveat: the exact parameter count is a moving target. Mistral's own listing has reported 675B total parameters with 41B active, while the announcement and model documentation put the figure near 1.05 trillion total with around 49–52B active. The architecture is a granular mixture of experts — a design that spreads activation across many smaller expert groups rather than a few huge ones — paired with a roughly 1.6-billion-parameter vision encoder. Training ran on about 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.

Where it lands against the field

Mistral's published preview results show 61.7 on DeepSWE v1.1 and 59.9 on AutomationBench, ahead of GLM 5.3, Qwen 3.8 Max, and Reflection's Beam, and behind DeepSeek V4.1 Flash and Kimi K3. Independent evaluation from Artificial Analysis gives it an Intelligence Index score of 38 — level with GPT-6 Luna at maximum reasoning effort — and a Cyber Index score of 50. On a broad aggregate index it ranks 32nd of 44 models, which is why the fair read is less 'new overall champion' and more 'best-in-class for what Mistral specializes in': independent evaluations back up the cybersecurity and legal-drafting claims, put finance in the middle of the pack, and rate coding, mathematics, and long-context work as mixed.

The preview API is live today at roughly $1.36 per million input tokens and $4.18 per million output tokens. Because the preview serves the same model as the final release, the practical move is to run your own prompts and evaluation sets against it now rather than waiting.

Open weights: the part that matters

The headline commitment is that the weights drop by the end of October — with October 27 circulating in secondary coverage, the French announcement naming the 26th, and a Hugging Face listing showing the 31st, so treat the exact date as unconfirmed. The license text is not yet final; Mistral calls the model open-weight and targets self-deployment on private cloud or on-premise, but unlike Mistral Large 3 (which shipped under Apache 2.0), the terms have not been published. The practical self-hosting bill is substantial: roughly 525 GB of weights at INT4, 1.05 TB at FP8, or 2.1 TB at BF16, making an eight-GPU Blackwell or H200 node the sensible minimum production config. MoE sparsity saves compute, not VRAM.

The sovereign-AI pitch runs parallel to the spec sheet. Mistral leads with data that has to stay on your own infrastructure, deployments that need to be auditable, and fine-tuning workloads — all served from models trained and run in European data centers, in every official EU language. Whether that sells depends on the eventual license, which is the missing piece right now.

• Test the preview API now — it is the same model the final release will serve
• Track the weights release and license text before committing to self-hosting
• Benchmark Large 4 against your own workloads, not leaderboards, before production use

What this means for you

If you build on frontier models, Large 4 adds a realistic European open-weight frontier option that you can prototype in the API today and potentially self-host within the month — a combination the closed labs do not offer. The open question is whether the final license and the independent benchmark picture on coding and general capability survive contact with the actual weights drop.

Sources: DataCamp — Mistral Large 4 (Le Chonk): Everything We Know; Yotta Labs — Mistral Large 4 Hardware Requirements (2026); CellCog — Mistral Large 4 (Le Chonk): Specs, Price, Benchmarks.

Corrections: Report an error at corrections@wireshaman.com.