Mistral Large 4 opens in preview with 1 trillion parameters

Mistral AI opened a public preview of Mistral Large 4 on October 6, 2026. Its launch post calls the 1-trillion-parameter model the company's largest yet and says the weights will be released by the end of the month.

Mistral is pitching Large 4 as a European option between closed US models and open models that often come from China, TechCrunch reported. Pierre Stock, Mistral's VP of Science, told TechCrunch it will work with "trusted partners and governments" before the release. He said the goal is weights that "can be used to defend," not to attack.

Mistral Large 4 has 1.05 trillion parameters and a 1M context

Large 4 is a mixture-of-experts model, so only a small part of it runs for each token. It combines instruct and reasoning modes in one model. It reads text and images but writes only text.

• Total parameters: 1.05 trillion, which the launch post rounds to 1 trillion
• Active parameters: 52 billion, plus a 1.6 billion-parameter vision encoder
• Context window: 1 million tokens in Mistral's docs, 524,000 in Artificial Analysis's listing
• Preview price: $1.36 per million input tokens, $0.14 cached, $4.18 output
• Training: from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers
• Languages: more than 160, including every official EU language
• Weights: due by the end of October

Source: Mistral's launch post and model docs, checked October 11.

Outside tests put Large 4 in the pack, not in front

In Mistral's own tests, Large 4 scores 61.7% on the DeepSWE coding benchmark and solves 93% of 40 Cybench security challenges. In a blind coding review run with Surge AI, it placed second of five models, behind Claude Opus 5.

Those are vendor-reported results, and outside checks are mixed. VentureBeat's benchmark check found the live DeepSWE leaderboard puts GLM-5.3 and Kimi K3 at about 69% in their best setups. That means 61.7% isn't a coding lead, though VentureBeat did confirm Mistral's rival figures on a legal benchmark against Vals.ai's public board.

Artificial Analysis scores the preview 38 on its Intelligence Index. That ranks it 64th of 227 models, above the median of 26 for comparable models. It's also wordy: it used 200 million output tokens to run the index, against a median of 82 million.

Mistral is betting on cyber work that closed models refuse

Mistral says Large 4 ranks in the top five of the Artificial Analysis Cyber Index. That index tests whether a model can find, reproduce and patch flaws in real code. Recaply couldn't confirm the rank, because the index's public page names only its top three.

On one of the index's tests, Mistral says Large 4 scores 82%, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse. The index counts a refused task as a zero. Mistral argues defenders need a model that won't block legitimate security research.

Until the weights ship, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get reduced moderation and expanded cyber skills. Mistral also says Large 4 refuses malicious cyber prompts more often than any other open model it tested.

Developers welcomed a European model but questioned its value

The Hacker News launch thread drew 1,211 comments. Recaply read the top 30 on October 11, a snapshot rather than a survey.

About four in ten welcomed it, often as a European or open-weight option. About a quarter called it underwhelming, often saying cheaper models such as DeepSeek 4.1 Flash and GLM-5.3 match it. One commenter asked whether its cyber strength makes it easier to use for attacks.

Very glad to see Mistral still in the game even after some big stumbles with Large 3.

eigenspace on Hacker News, October 6

Large 4 is the first model from Mistral's €3B round

Mistral calls Large 4 the first milestone funded by its €3B Series D round. It says that's the largest equity round ever raised by a European tech company. Samsung led the round in September at a €21B valuation, according to TechCrunch.

Large 3, the previous flagship, had 675 billion parameters with 41 billion active, VentureBeat noted. Mistral plans to use Large 4 as the base for a new set of specialized models.

The preview API is live on Mistral Studio, served from Mistral's own data centers in Europe. Mistral says the model will also run in other regions, including a European deployment it operates under European law.

The weights, architecture details and more benchmarks are due by the end of October. VentureBeat reported a target date of October 27, under a custom Mistral license.