Mistral AI has opened a public preview of Mistral Large 4, a multimodal model with 1 trillion total parameters, and plans to publish its weights later this month once safety testing is complete, SiliconANGLE reported Tuesday. The Paris-based company is positioning the model as the strongest open-weight system built outside China.

What Happened

Large 4 uses a mixture-of-experts design. Of its 1 trillion parameters, only about 49 billion are activated for any given prompt, which SiliconANGLE noted makes it far less demanding on hardware than running the whole network. The model can respond in more than 160 languages. TechCrunch reported that Mistral staff have nicknamed it "Le Chonk."

For now, access runs through a guarded public endpoint on Mistral's own cloud platform. TechCrunch said the open-weight release is planned for roughly three weeks out, after safety testing. Pierre Stock, Mistral's vice president of science, told TechCrunch the company will work with trusted partners and governments so that the released weights help defenders rather than attackers.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

The two outlets described the training hardware differently. SiliconANGLE reported the run used 3,800 Grace Blackwell superchips, each pairing two Nvidia Blackwell GPUs with a CPU. Stock told TechCrunch the model was trained on about 4,000 Nvidia GPUs, which he said was two to three times fewer than Chinese competitors use.

How It Scores

The benchmark results Mistral published are mixed. According to SiliconANGLE, Large 4 ranked in the top five on the AA Cyber Index, a set of tests of whether models can find and fix software vulnerabilities, and its 82% score on open-source patching put it ahead of other open models. On Dense200, which measures how well a model locates objects in images, it edged OpenAI's GPT-6 Astra by 1%.

Coding is the weak spot. SiliconANGLE reported that Large 4 falls well behind closed frontier models such as Astra on the most widely used coding benchmarks, although it beats the open models Qwen3.8 Max and DeepSeek V4 Pro. It also outscored DeepSeek V4 Pro on AutomationBench, made up of relatively simple tasks, and on AA-Briefcase, which includes assignments that would take a person weeks. TechCrunch noted that independent benchmark results were still pending at launch.

Why It Matters

Mistral is one of Europe's best-funded AI developers, and Large 4 is its bid to show a European lab can compete at the trillion-parameter scale. TechCrunch reported that the company frames open-weight but guarded releases as a third way between closed American models and open Chinese ones, with Stock arguing that open weights are easier to audit.

The chosen markets track Mistral's investors. Stock told TechCrunch the model is tuned for cybersecurity and finance as well as chip design. Dutch lithography equipment maker ASML led Mistral's Series C in 2025, and Samsung led its Series D in September at a valuation of €21 billion, about $24.4 billion.

Training Setup

SiliconANGLE described a reinforcement-learning pipeline that runs tens of thousands of parallel rollouts drawn from a modular library of environments, including code sandboxes, a web search engine and tests that check results. Rollouts and training run asynchronously to keep the hardware busy, and the rollouts produced 33 billion tokens a day. Mistral did not stop that training run after producing the current version.

What to Watch

The next milestone is the weights release later in October, including whether Mistral limits access to the most security-sensitive capabilities. The company also said it will build a series of models tuned for specific use cases on top of Large 4, and it expects larger, more capable versions to come out of the continuing training run in the coming months. Independent testing will show whether the cybersecurity and vision results hold up outside Mistral's own evaluations.