The French company Mistral AI has unveiled Mistral Large 4, a multimodal model featuring 1.05 trillion parameters and open weights. The firm positions it as a competitor to leading American and Chinese neural networks, focusing on programming, cybersecurity, and enterprise applications.
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.
— Mistral AI (@MistralAI) October 6, 2026
Mistral Large 4 is built on a Mixture of Experts (MoE) architecture. While it boasts 1.05 trillion parameters, only about 49 billion are active during request processing. Additionally, the model includes a separate visual encoder with 1.6 billion parameters and supports a context of up to 1 million tokens.
The company asserts that Large 4 ranks among the most efficient open-weight models and significantly narrows the gap with closed systems. Mistral also claims that this new model outperforms some Chinese open solutions, particularly in cybersecurity tasks. However, the company has not provided independent results to back these claims.
Source: Mistral.Training on 4,000 Nvidia Accelerators
Mistral trained Large 4 independently over approximately two months using its infrastructure in European data centers, utilizing around 4,000 NVIDIA Grace Blackwell GPUs.
The model is designed for a broad range of tasks, including programming, cybersecurity, finance, geospatial analysis, manufacturing, and product design. According to the company, Large 4 can also assist in computer chip development.
Mistral emphasizes its focus on agent-based scenarios and tasks that require processing large volumes of information and executing sequential actions.
Open Weights vs. Closed Models
A key distinction of Large 4 compared to models from OpenAI and Anthropic is its open weights. After the release, developers will be able to run the model independently and tailor it to their needs without sharing data with an external closed AI provider.
Currently, Large 4 is available in public preview. The full release of the weights is scheduled for October 27. Until then, Mistral is offering early access to developers, cybersecurity experts, and government organizations for testing the model.
The company is also evaluating the model's ability to operate beyond its specified environment. Mistral representatives noted that during testing, Large 4 attempted such actions, but software limitations prevented them.
A European Response to OpenAI, Anthropic, and Chinese Models
The launch of Large 4 comes amid a renewed competition in the open-weight model segment. Chinese companies like Alibaba, Z.ai, Moonshot, and DeepSeek have recently strengthened their positions, while Western developers have traditionally led in closed models.
Now, Mistral aims to position itself in an intermediary role: offering a model on par with advanced systems while allowing users control over its weights and infrastructure, as noted by TechCrunch.
The company views this release as a significant milestone following a recent funding round of €3 billion, which raised Mistral's valuation to over $24 billion. Among its new investors is Samsung Electronics.
It’s worth noting that on October 5, the Nvidia-backed American startup Reflection AI released its first open-weight model with 501 billion parameters.
Follow ForkLog on social media
Telegram (main channel) Facebook X