Mistral Large 4 preview: a 1-trillion-parameter open-weight model, weights due this month

Mistral AI opened a public preview of Mistral Large 4 on October 6, 2026. It is the Paris company's biggest model yet: a mixture-of-experts (MoE) model with about 1 trillion total parameters and 52 billion active per token, nicknamed "le Chonk". Right now you can only use it through Mistral's API. Mistral says the downloadable weights will follow by the end of October, after a few weeks of testing with security researchers, partners and government authorities.

What Mistral announced

Mistral calls Large 4 its most capable model yet, and "at the frontier of open weight models". Here is what the announcement and model docs confirm:

For comparison, Mistral Large 3 had 675 billion total and 41 billion active parameters, so this is a big jump in size.

Benchmarks (Mistral's numbers)

All of these come from Mistral and haven't been checked independently. When VentureBeat went to press, the model wasn't yet listed on Artificial Analysis or the public DeepSWE leaderboard.

Cybersecurity is the headline. Mistral says Large 4 is in the top five on the Artificial Analysis Cyber Index. It says closed models such as Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduce-and-patch test because they refuse the task. That says as much about safety policies as about raw skill, so read it with care.

On agentic coding, the picture is less rosy. VentureBeat notes that on the live DeepSWE leaderboard, GLM-5.3 and Kimi K3 sit around 69%, and the top closed models around 74%. Large 4's 61.7% is good for an open model, but it isn't the best open model on that test.

Pricing and access

The preview is live in Mistral Studio (console.mistral.ai), and the docs page is mistral-large-4-0. Check the exact model ID in the API docs before you hard-code it.

Mistral's docs list function calling, structured outputs and batching as supported. Enterprises can also deploy it privately or on-premise, and Mistral is pitching zero-data-retention and "sovereign" European hosting for regulated industries.

The open-weights question

The weights are the whole reason this release matters. A trillion-parameter model with open weights from a European lab would give companies a serious option they can host themselves, outside both the US and China. But some things are still unknown:

• License: Mistral hasn't said. VentureBeat reports a custom Mistral license is expected, so don't assume Apache 2.0.
• Date: the official line is "by the end of October". VentureBeat reported a target of October 27.
• Hardware: even with only 52B active parameters, you need enough memory to hold all ~1T parameters. Running it yourself is a job for multi-GPU servers, not a workstation, unless smaller or quantized versions appear.

Mistral says the preview period is for red-teaming with cybersecurity leaders, vetted partners and state authorities, who get reduced moderation for that work. It also reports strong scores on prompt-injection and jailbreak resistance benchmarks, but again, these are its own numbers.

Who should care

• European companies with data-residency rules: this is the most capable model you can run in the EU on European infrastructure, and soon on your own hardware.
• Security teams: if Mistral's numbers hold up, Large 4 will do defensive security work that closed models often refuse.
• Builders on a budget: at the current discounted preview price, it's one of the cheapest large models to test through an API.
• Open-model fans: wait for the weights and independent benchmarks before you switch from GLM, Kimi, Qwen or DeepSeek.

Bottom line

Mistral Large 4 is a big, credible step for open-weight AI outside China, but today it's still a preview API. If you're curious, test it in Mistral Studio while the discount lasts, on your own tasks. If you plan to self-host, wait until the end of October for the weights, the license terms and independent benchmarks before you commit. We'll cover the weights release when it lands.