Mistral Large 4: open weights due in late October
Mistral's new flagship packs a trillion parameters, 49 billion of them active. For now it runs only through the API; the weights arrive in late October.
In short
Mistral Large 4 is a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49 billion active per token, in API preview since October 6, 2026, with Mistral saying the open weights follow at the end of October 2026.
At a glance
- 1 trillion parameters in total, 49 billion active per token; inputs are text and image.
- Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
- Preview pricing: $1.36 per million input tokens, $4.18 per million output tokens.
- Artificial Analysis score of 38. Mistral Large 3 from December 2025 scored 9 on the same index.
- Code: 61.7% on DeepSWE v1.1, 28.3% on Terminal Bench 4.0; 93% on Cybench.
Mistral Large 4 is a natively multimodal mixture-of-experts model carrying 1 trillion parameters, 49 billion of which are active per token; it went into API preview on October 6, 2026, and Mistral says the open weights ship at the end of October 2026.
One model, two thinking levels
Mistral describes ML4 as a hybrid that folds instruction following and reasoning into one network instead of shipping them as separate checkpoints. The API exposes exactly two reasoning levels, none and high, so the caller decides how much the model deliberates. It accepts text and images, and Mistral claims more than 160 languages including every official EU language.
The Le Chonk nickname comes from the announcement itself, as developer Simon Willison read the post's title. It is a joke about sheer size. The reference point is Mistral Large 3 from December 2025, which managed 9 points on the Artificial Analysis index.
Reading the scores
On code, Mistral reports 61.7% on DeepSWE v1.1, 28.3% on Terminal Bench 4.0, 59.4% on SWE-Atlas-QnA and 49.8% on its own Coding Agent Index. A human preference study run by Surge AI put ML4 at 3.74 out of 5.0, second among five models tested.
Security work is where the vendor's numbers run highest: 93% on Cybench, 82% on a vulnerability reproduce-and-patch test, and 93.3% attack resistance on B3. For multimodal, Mistral lists 42% on Dense 200 against 41% for GPT-6-Astra.
Independent scoring is cooler. Artificial Analysis places ML4 at 38, just behind DeepSeek 4.1 Flash — a 552-billion-parameter model, roughly half the size. Willison's read is that this is not a Fable-class model and that Mistral sits about six months off the frontier, which is still a long way up from Large 3's score of 9.
Price, and where it was trained
Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the training run used 3,800 NVIDIA Grace Blackwell GPUs in datacenters it operates in Europe.
For a company selling itself as the European option, where the run happened is part of the pitch — and it is an easier claim to check than any benchmark table.
What we could not verify
Neither source we read names the license the weights will carry. So "open weights" is Mistral's wording at this stage, not a verified licensing fact.
A third trade source (marktechpost.com) returned HTTP 403 to our request. Its headline figure of 1.05 trillion parameters is therefore unchecked, and we stay with Mistral's own figure of 1 trillion. We also have no dependable number for the hardware a model this size needs to run locally.
FAQ
How many parameters does Mistral Large 4 have?
1 trillion in total, with 49 billion active per token. The figure comes from Mistral's own announcement.
When do the Mistral Large 4 open weights come out?
Mistral says the end of October 2026. The license they will carry is not named in the sources we read.
What does Mistral Large 4 cost on the API?
Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens.
