Technical Specifications and Architecture

Mistral AI has announced the release of Mistral Large 4, a 1 trillion-parameter natively multimodal Mixture-of-Experts (MoE) model. According to company documentation, the model utilizes a granular architecture featuring 128 routed experts, with 4 active parameters per token. This results in approximately 49 billion active parameters per token, or 52 billion when including embeddings and output layers.

The model was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs located in Mistral’s European datacenters. It features a 1-million-token context window designed for long-context workloads and utilizes a Pixtral-style vision encoder capable of processing variable image sizes and aspect ratios.

Capabilities and Performance Benchmarks

Mistral Large 4 is designed to unify instruction following, reasoning, coding, and agentic workflows. The company has introduced a “Reasoning Mode” that allows users to configure reasoning effort across three levels: low, medium, and high.

Key functional capabilities include:

  • Multilingualism: Native fluency in over 160 languages, including all official languages of the European Union.
  • Vision: The ability to analyze technical drawings, charts, documents, and images, with the company claiming state-of-the-art visual grounding among open-weight models.
  • Agentic Workflows: Support for native function calling, JSON output, and strict JSON schema enforcement.
  • Specialized Domains: The model is optimized for workloads in cybersecurity, finance, and law. Mistral claims the model is state-of-the-art among open-weight models for security tasks, while maintaining protocols to refuse malicious requests.

Licensing and Availability

The model is being released under the Mistral Ecosystem Open-Weight License, a modified version of the MIT license. This allows for both commercial and non-commercial use, though Mistral specifies that entities operating a Model-as-a-Service business require a separate commercial license.

For developers, the model is available via the Mistral API under the identifier mistral-large-4.

Industry Reception and Discussion

Following the announcement, discussions on platforms such as Reddit have highlighted various technical and strategic implications. Some users questioned the utility of releasing weights for a model of this scale, suggesting the primary beneficiaries may be competing AI firms. Others noted the significant hardware requirements for local deployment; unverified comments on social media suggest that running such a model may require substantial VRAM, with some estimates suggesting requirements as high as 400GB to 600GB depending on the timeframe and hardware configuration.