Mistral Launches Large 4 Preview, a 1 Trillion-Parameter AI Model

Mistral AI launched a public preview of Large 4 on October 6, a 1 trillion-parameter native multimodal mixture-of-experts model with 49 billion active parameters, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers. The preview is priced at $1.36 per million input tokens and $4.18 per million output tokens, with downloadable weights planned for later in October after red-teaming with cybersecurity firms, partners and government agencies. Artificial Analysis data cited by The Decoder gave the preview 38 points on its Intelligence Index, versus 9 for Mistral Large 3 and 58 for Claude Opus 5.5, and Mistral said Large 4 scored 82% on a vulnerability-reproduction and patching test while some closed models refused the task.

Mistral Launches Large 4 Preview, a 1 Trillion-Parameter AI Model - Mistral Large 4 has 1 trillion total parameters and 49 billion active parameters. <sup \ 1\ https://mistral.ai/news/mistral-large-4/ </sup - The model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers. <sup \ 1\ https://mistral.ai/news/mistral-large-4/ </sup - Mistral plans to release the weights later in October after red-teaming with cybersecurity firms, partners and government agencies. <sup \ 1\ https://mistral.ai/news/mistral-large-4/ </sup <sup \ 2\ https://www.axios.com/2026/10/06/reflection-mistral-open-weight-ai-models-china </sup <sup \ 4\ https://www.lemonde.fr/en/economy/article/2026/10/06/mistral-ai-unveils-new-ai-model-aimed-at-narrowing-the-gap-with-top-chinese-competitors 6758318 19.html </sup - Mistral says Large 4 scored 82% on a vulnerability-reproduction and patching test, while some closed models refused the task. <sup \ 1\ https://mistral.ai/news/mistral-large-4/ </sup <sup \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ </sup - Artificial Analysis data cited by The Decoder puts Large 4 well above Mistral’s previous models but below the leading closed models and several Chinese open-weight systems. <sup \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ </sup Mistral AI on October 6 launched a public preview of Large 4, a 1 trillion-parameter multimodal model that the French company describes as its largest and most capable system to date. The model has 49 billion active parameters and is available through Mistral’s API, with downloadable weights planned for later in October. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 2\ https://www.axios.com/2026/10/06/reflection-mistral-open-weight-ai-models-china \ 4\ https://www.lemonde.fr/en/economy/article/2026/10/06/mistral-ai-unveils-new-ai-model-aimed-at-narrowing-the-gap-with-top-chinese-competitors 6758318 19.html Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers. Mistral said the model is designed for coding, agentic workflows, multimodal analysis and enterprise applications including cybersecurity, finance and law. \ 1\ https://mistral.ai/news/mistral-large-4/ The release gives Mistral a larger prospective open-weight offering as Chinese companies lead much of the current open-model market and US companies retain the strongest closed systems. Artificial Analysis data cited by The Decoder gave the preview 38 points on its Intelligence Index, compared with 9 for Mistral Large 3 and 58 for Claude Opus 5.5. \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ The Model Mistral describes Large 4 as a native multimodal mixture-of-experts model that combines instruction following, reasoning and agentic capabilities. The company said it supports more than 160 languages, including every official language of the European Union, and is trained for tasks involving documents, charts, technical drawings and satellite imagery. \ 1\ https://mistral.ai/news/mistral-large-4/ The preview is priced at $1.36 per million input tokens and $4.18 per million output tokens, according to Mistral’s announcement. The company plans to publish additional architecture, licensing and post-training details when the weights are released. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ Because the weights have not yet been published, Large 4 is currently available as a hosted preview rather than a downloadable open-weight model. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ Cybersecurity as the Differentiator Mistral made cybersecurity the central focus of its launch materials. On the Artificial Analysis Cyber Index, the company said Large 4 ranks among the five strongest models globally and leads open-weight models developed outside China. It reported an 82% score on a test requiring the model to reproduce a real software vulnerability and then patch it, along with a 93% score on the Cybench security benchmark. \ 1\ https://mistral.ai/news/mistral-large-4/ Mistral said Claude Opus 5.5 and GPT-6 Astra scored near zero on the vulnerability task because they refused to perform it. The company is using that distinction to argue that self-deployment can give security teams more control over legitimate vulnerability research and incident response. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ The model is not being released without safeguards. Until the weights are published, Mistral is testing a version with reduced moderation and expanded cyber capabilities with cybersecurity leaders, vetted partners and state authorities. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 2\ https://www.axios.com/2026/10/06/reflection-mistral-open-weight-ai-models-china Performance and Competition Large 4 represents a substantial improvement over Mistral’s recent models on the Artificial Analysis Intelligence Index, according to The Decoder’s review of the benchmark. Its score of 38 exceeded the reported scores for Mistral Large 3 and Mistral Medium 3.5, but remained below leading models from Anthropic, OpenAI and Google as well as several Chinese open-weight systems. \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ Mistral reported a 49.8% score on the Artificial Analysis Coding Agent Index and a 59.9% score on AutomationBench. In a blind coding evaluation conducted with Surge AI, Large 4 ranked second among five models, behind Claude Opus 5. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ Le Monde reported that Mistral characterized the model as roughly comparable with leading Chinese models from one or two months earlier, while noting that the company’s benchmark figures remain preliminary and will be expanded when the weights are released. \ 4\ https://www.lemonde.fr/en/economy/article/2026/10/06/mistral-ai-unveils-new-ai-model-aimed-at-narrowing-the-gap-with-top-chinese-competitors 6758318 19.html European Infrastructure and Open Weights Mistral said the model and its public preview run on infrastructure it operates in Europe, including a European deployment managed independently of other digital-service providers and under European law. The company is framing that setup as a way for customers to retain control over data, deployment and model access. \ 1\ https://mistral.ai/news/mistral-large-4/ Axios reported that Mistral plans to release the weights on October 27 after additional reinforcement learning and safety testing. Once released, the weights are expected to give customers the ability to deploy and customize the model outside Mistral’s hosted API, although the final license and technical documentation had not been published at launch. \ 2\ https://www.axios.com/2026/10/06/reflection-mistral-open-weight-ai-models-china \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ Mistral’s release follows a €3 billion Series D announced in September. The company has been expanding its European computing infrastructure as it seeks to compete with better-funded US labs and rapidly advancing Chinese model developers. \ 1\ https://mistral.ai/news/mistral-large-4/ \ 3\ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ \ 4\ https://www.lemonde.fr/en/economy/article/2026/10/06/mistral-ai-unveils-new-ai-model-aimed-at-narrowing-the-gap-with-top-chinese-competitors 6758318 19.html Companies mentioned Further sources \ 1\ Mistral AI, “Introducing Mistral Large 4,” October 6, 2026. ↗ https://mistral.ai/news/mistral-large-4/ \ 2\ Axios, “Western AI labs challenge China’s open-model lead,” October 6, 2026. ↗ https://www.axios.com/2026/10/06/reflection-mistral-open-weight-ai-models-china \ 3\ The Decoder, “Mistral Large 4 is Europe’s trillion-parameter answer to US model… ↗ https://the-decoder.com/mistral-large-4-is-said-to-be-the-most-powerful-open-ai-model-from-europe-and-the-u-s/ \ 4\ Le Monde, “Mistral AI unveils new AI model aimed at ‘narrowing the gap’ with to… ↗ https://www.lemonde.fr/en/economy/article/2026/10/06/mistral-ai-unveils-new-ai-model-aimed-at-narrowing-the-gap-with-top-chinese-competitors 6758318 19.html The stories that matter, in one email. Free — unsubscribe anytime.