Mistral Large 4 launches with 1.05T parameters

Mistral AI released Mistral Large 4 in public preview on October 6, 2026, a mixture-of-experts model with 1.05 trillion parameters and a 1 million token context window. It is available now through Mistral's API, and Mistral's announcement promises the open weights "by the end of the month." That makes it one of the largest models a team will be able to download and run on its own servers.

What Mistral released

Mistral Large 4 is a general-purpose model that reads both text and images. Mistral calls it "natively multimodal," meaning images are part of the model from the start, not a separate add-on. Its nickname is "Le Chonk."

It is a mixture-of-experts model, or MoE. An MoE model splits its parameters into many smaller "expert" networks. For each token, a router picks only a few experts to do the work. So the model stores a huge number of parameters but uses only a small share of them at a time.

Mistral's own pages disagree on the exact sizes:

Spec	Announcement	Model card
Total parameters	"1 trillion"	1.05T
Active parameters per token	49 billion	52B
Vision encoder	not stated	1.6B
Context window	1 million tokens	1M tokens

Unite.AI and The Bright Minded both report 1.05 trillion total and 49 billion active. Until Mistral corrects one page, treat the active count as somewhere between 49 and 52 billion.

The model supports more than 160 languages, including every official language of the European Union, according to Unite.AI. Mistral says it trained the model on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers.

Price and availability

The API model name is mistral-large-4. Mistral's announcement lists $1.36 per million input tokens and $4.18 per million output tokens.

Unite.AI reported different numbers: $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. Mistral's own pages do not mention those prices or a cached-input rate. Check the price in your Mistral console before you budget against either figure.

The open weights are the other half of the launch. Mistral says it will release them by the end of October. The Bright Minded cites an October 27 date from Reuters reporting. Neither the announcement nor the model card names the license the weights will ship under.

Benchmarks and the security angle

Mistral published its own benchmark results. These are the company's numbers, not independent tests.

Benchmark	What it tests	Mistral Large 4
DeepSWE v1.1	Software engineering tasks	61.7%
SWE-Atlas-QnA	Questions about codebases	59.4%
AutomationBench	Task automation	59.9%
Terminal Bench 4.0	Command-line tasks	28.3%
Cybench	Security challenges	93%
AA Cyber Index	Reproducing software vulnerabilities	82%

On human ratings, Mistral scores Large 4 at 3.74 out of 5. Claude Opus 5 scored 4.22, GLM-5.3 scored 3.60 and Kimi K3 scored 3.59 in the same comparison.

The cyber score is the headline. Unite.AI says 82% puts the model in the global top five on the Artificial Analysis Cyber Index. It is also the top open-weight model from outside China on that index.

Mistral is handling that capability with two versions. Its announcement says it is red-teaming the model "with cybersecurity leaders, vetted partners, and state authorities." Those groups "will access the same model with reduced moderation and expanded cyber capabilities." Everyone else gets the standard, more filtered model.

Mistral has grown fast this year. It raised €3 billion at a €21 billion valuation in September. The launch also lands a day after Reflection AI announced Beam, a 501B open-weight model, whose weights are also due later in October.

What this means for developers

• Try the preview now, plan production later. The API works today under mistral-large-4. Run your own evaluation set through it while the price is known and the weights are not.
• Wait for the license before you plan self-hosting. "Open weights" can mean many license terms. Do not promise a customer on-premises deployment until Mistral publishes them.
• Size your hardware for the total, not the active count. Only about 50 billion parameters work on each token, but all 1.05 trillion must sit in memory. Self-hosting means a multi-GPU server, not a single card.
• Settle the spec and price questions in writing. The 49B versus 52B gap and the two price lists will matter for capacity and cost planning. Ask Mistral support, or wait for the release model card.
• Security teams should look closely. A model this strong at reproducing vulnerabilities helps defenders with triage and testing. It also means your threat models should assume attackers can run something similar once the weights ship.

The weight release is the date to watch. Until then, Mistral Large 4 is a hosted API, and its open-weight promise is still a plan.