Mistral Large 4 Opens 1T-Parameter Preview Before Open Weights

Mistral AI opened a public preview of Mistral Large 4 on Tuesday, 6 October 2026, a natively multimodal model with 1 trillion parameters that the Paris company says it will release as open weights by the end of the month. In its launch post, Mistral said the model, which it calls ML4 and nicknames "le Chonk," runs 49 billion active parameters and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. Developers can call the preview API on Mistral Studio now, while cybersecurity firms, vetted partners and state authorities test a less restricted version before the weights ship.

The launch is Mistral's first new model since May, Reuters reported, and Mistral describes it as the first milestone on the roadmap funded by its €3 billion Series D. Reuters also reported that the company plans to make the model fully public on 27 October.

What Mistral released on Tuesday

The preview is an API product first. Mistral's model documentation lists Mistral Large 4 as a public preview, version v26.10, with a granular mixture-of-experts design, 1.05 trillion total parameters, 52 billion active parameters, a 1.6 billion parameter vision encoder and a 1 million token context window. The documentation's parameter figures are slightly higher than the rounded numbers in the launch post, and Mistral has said it will publish fuller architecture details when the weights arrive.

The documentation also lists structured outputs, function calling, document question answering, batching, and the Agents and Conversations endpoints among the supported features. The launch post says a significant share of the training data was multilingual, covering more than 160 languages, including every official language of the European Union.

TechCrunch reported that Mistral VP of Science Pierre Stock put the training fleet at 4,000 Nvidia GPUs, "which is two to three times less than our Chinese competitors," a rounder figure than the 3,800 in the company's post. TechCrunch also noted that benchmark results were still pending at the time of its interview, and that Stock named cybersecurity, finance and chip design among the model's target uses.

The benchmark claims behind the launch

Most of the performance numbers come from Mistral itself, and the company says they will change as training continues. On the Artificial Analysis Cyber Index, Mistral says ML4 ranks among the top five models globally. On one test inside that index, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral reports a score of 82 percent, which it says is the highest of any model. It also reports solving 93 percent of the 40 challenges in Cybench.

Mistral argues that some closed models score near zero on that reproduce-and-patch test because they refuse the task, naming Claude Opus 5.5 and GPT-6 Astra. SiliconANGLE also highlighted the 82 percent figure and the top-five Cyber Index placement in its coverage.

For software work, the launch post lists 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4, for a combined Coding Agent Index score of 49.8 percent. In a blind human evaluation run with Surge AI, annotators rated coding outputs on a one-to-five scale and placed ML4 Preview second of five models at 3.74, behind Claude Opus 5 at 4.22. On AutomationBench, a set of 657 business workflows across apps such as Gmail, Google Sheets, Slack and Salesforce, Mistral reports 59.9 percent.

On safety, Mistral says ML4 resisted 93.3 percent of attacks on Lakera's B3 AI Security Benchmark, and that its average refusal rate on malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm is higher than all the open-source models it compared.

Why the weights are arriving later

Mistral is splitting the launch into two stages. Until the weights are published, the company says it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.

"In the meantime, we'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks," Stock told TechCrunch. Reuters reported that, during the preview period, some cybersecurity experts and state authorities would have access to a version with fewer safety barriers to test its capabilities.

The company also pitches location and control. The launch post says the model will be available in multiple regions, including a European deployment that Mistral operates end to end under European law, and that the preview runs on the same infrastructure used to train it. According to TechCrunch, Mistral's backers include ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation. Reuters reported that Stock said Large 4 is capable of designing computer chips but declined to say whether Mistral is developing chips with Samsung.

What comes before the October release

Mistral says the reinforcement learning run behind the preview is still in progress and shows no signs of saturation. At its current scale of roughly 3,000 GPUs, the company says one training run produces about 33 billion tokens per day, around 16 billion of which are trainable completion tokens after filtering.

Alongside the weights, Mistral has promised more detail on the architecture, additional benchmarks and its post-training methodology. It also says ML4 will serve as the base for a new set of specialized models built for particular industries and workloads.

Other open-weight developers have also been active in recent weeks. PromptCrates recently covered Reflection's Beam open-weight model, Z.ai's confirmation that it built Ox Alpha, and Anthropic's warning about GLM-5.3's cyber capabilities, the same model family Mistral now uses as a comparison point in its human evaluations. Whether the self-reported scores hold up should become clearer once outside evaluators can run the published weights.

• Mistral AI: Introducing Mistral Large 4
• Mistral Docs: Mistral Large 4 model page
• TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals
• Reuters via Euronext: France's Mistral announces new AI model
• SiliconANGLE: Mistral launches Mistral Large 4, details AI roadmap
• PromptCrates: Reflection debuts Beam open-weight model
• PromptCrates: Z.ai confirms it built Ox Alpha
• PromptCrates: Anthropic warns about GLM-5.3 cyber capabilities