Mistral ships le Chonk, a 1T open model. Europe has a contender, not a leader.

Mistral released Mistral Large 4 as a public preview on October 6. The company's own announcement says it is "unofficially ML4, very officially: le Chonk." It is Mistral's largest model, its first major release in about five months, and the first product of a €3 billion Series D at a €21 billion valuation, which CNBC reports closed in September. We have added it to our model database.

The spec sheet

Mistral describes ML4 as a natively multimodal Mixture-of-Experts model with about 1 trillion total parameters and 49 billion active per token. Its own docs card says 1.05 trillion and 52 billion, so treat both as approximate until the weights ship. It reads text and images and writes text. Mistral says it was trained on its own European infrastructure across more than 160 languages, including every official EU language.

The weights are the part that matters for Mistral's pitch, and they are not out yet. The announcement promises them "by the end of the month"; CNBC and Sifted report October 27. The license has not been published.

	Mistral Large 4
Parameters	~1T total, ~49B active (MoE)
API price per 1M tokens	$1.36 input / $4.18 output ($0.14 cached input)
Context window	1M per Mistral docs (512K on OpenRouter)
API model ID	mistral-large-4-0
Open weights	Promised for late October; license not yet announced

How good is it?

Mistral's own numbers are respectable rather than leading: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and 59.9% on AutomationBench. Its strongest claims are in security, with 93% on Cybench and 82% on a vulnerability reproduce-and-patch test where, Mistral says, Claude Opus 5.5 and GPT-6 Astra score near zero. That comparison mostly measures refusals: the closed models decline the task, so it says more about policy than about capability.

Independent testing is clearer. Artificial Analysis gives ML4 a score of 38 on its Intelligence Index, the highest of any model built outside the US and China. For scale, Claude Opus 5.5 scores 58 and Gemini 4 Argon 53. Among open-weight models, ML4 still trails several Chinese releases, including GLM-5.3 and Kimi K3.

"The model capabilities will further improve as we scale up our training capacity, following our Series D fundraise." - Guillaume Lample, Mistral co-founder and chief scientist, to CNBC

The price is the pitch, with a catch

At $1.36 in and $4.18 out, ML4 costs a fraction of the closed flagships per token. Mistral's docs also list a rate 50% lower that Artificial Analysis describes as a launch discount. Per task, though, cheap Chinese models still undercut it: Artificial Analysis puts ML4 at $1.13 per index task against $0.25 for GLM-5.3-Flash and $0.27 for DeepSeek's V4.1 Flash. If raw price-to-quality is your only criterion, ML4 is not the answer. Our AI Pricing Index tracks all three.

Who should switch

The buyer ML4 is built for is a European organization that needs to keep data and inference inside the EU, or wants to self-host a frontier-class open model on its own hardware. For that buyer there was no comparable option a month ago. Mistral pitches private-cloud and on-premises deployment directly, and its API is served from European infrastructure.

That buyer also has a second option now. Aleph Alpha released Kolibri the same week: a far smaller German-English model (78 billion total parameters, about 3 billion active) under the Apache 2.0 license, with weights already on Hugging Face. Sifted sums up the reaction as Europe now having two contenders in the open-weight race.

If you do not have a sovereignty requirement, the practical answer is unchanged. For the strongest coding and reasoning, the closed models in our model index still lead by a wide margin. For the cheapest capable open model, look at the Chinese releases. ML4 earns its place in the middle, and the late-October weights release will decide how much that matters.