Mistral Large 4 Arrives With 1.05 Trillion Parameters and Open Weights Coming This Month

Mistral has launched Mistral Large 4, its biggest and most capable model to date, in public preview.

The new model uses a Mixture-of-Experts architecture with 1.05 trillion total parameters and 49 billion active parameters, according to Mistral's model documentation. It is also natively multimodal, includes a 1.6 billion-parameter vision encoder, and supports a 1 million-token context window.

Mistral says the preview API is available now through Mistral Studio, while downloadable model weights are planned for later this month.

Read Mistral's official announcement

What Mistral Large 4 Is

Mistral Large 4 is a general-purpose multimodal model designed to handle text, images, reasoning, coding, tool use, and agent-style workflows.

Unlike a dense model where every parameter is active for every token, ML4 uses a Mixture-of-Experts, or MoE, design. Mistral lists 1.05 trillion total parameters, but only 49 billion are active during a given forward pass.

That architecture is intended to give the model much larger overall capacity while keeping inference more efficient than activating the full trillion-plus parameter count every time.

Mistral describes ML4 as an open-weight model, although the actual downloadable weights are not available yet during the initial public-preview period.

The Key Specs

Mistral's official documentation lists the following headline specifications:

• 1.05 trillion total parameters
• 49 billion active parameters
• 1.6 billion-parameter vision encoder
• 1 million-token context window
• multimodal input
• Mixture-of-Experts architecture
• support for structured outputs, function calling, document Q&A, batching, agents, and built-in tools

The model is currently listed as Public Preview v26.10.

That preview label matters because Mistral's lifecycle policy says public-preview models can still receive updates before reaching general availability.

View the Mistral Large 4 model card

Coding and Agentic Workflows Are a Major Focus

Mistral is positioning Large 4 heavily around coding and agentic work.

In its launch announcement, the company highlights performance across coding benchmarks, terminal-based software tasks, cybersecurity, finance, legal workflows, multimodal grounding, and autonomous-agent scenarios.

Mistral also says ML4 combines instruction-following, reasoning, coding, multimodal understanding, and agent capabilities inside one model.

That makes the model especially relevant to developers building AI systems that need to do more than answer a single prompt.

Examples include:

• coding assistants that can work through multi-step software tasks;
• agents that call tools and APIs;
• systems that analyze documents and images together;
• research workflows that combine retrieval, reasoning, and structured outputs; and
• enterprise automation involving specialized data.

Open Weights Are Coming Later This Month

One of the biggest parts of the announcement is the planned release of the model weights.

Mistral says the weights will be released by the end of October.

That is important because open-weight availability can let organizations run models on their own infrastructure, fine-tune them for specialized use cases, or deploy them in environments where they want more control over data and compute.

During the public preview, however, the main way to try ML4 is through Mistral's API and Studio.

Mistral says additional architecture details, benchmarks, and post-training information will arrive alongside the weight release.

Built in Europe

Mistral is also emphasizing where the model was trained.

The company says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers.

The public-preview API is being served from the same infrastructure.

Mistral frames this as part of its broader push for European AI sovereignty, particularly for governments and companies that want more control over where models are trained and deployed.

Mistral Says Training Is Still Ongoing

This is not being presented as a finished model.

Mistral says the reinforcement-learning run behind the current preview is still in progress and that the model is continuing to improve.

The company expects additional gains as it expands its compute capacity and continues post-training work.

That means benchmark numbers published during the preview may change before the final weight release.

It also means buyers and developers should treat ML4's current performance as a moving target rather than a fixed final result.

What Public Preview Means

Mistral's own lifecycle documentation says a public-preview model is not the same as a general-availability model.

Public-preview models can receive silent updates and do not have a guaranteed path to general availability.

For developers, that means the model is available to test now, but production deployments should still account for the possibility that behavior, performance, pricing, or other details may change.

Mistral's changelog also lists launch pricing at 50% off for two weeks, which makes the preview period a cheaper time to experiment with the API.

See Mistral's model changelog

Why Mistral Large 4 Matters

The most important part of this release is not simply that the model crosses the trillion-parameter mark.

ML4 combines that scale with open-weight plans, multimodal input, a long context window, and a strong focus on coding and agents.

That puts it directly into the growing competition around models that are designed to act as more complete work systems rather than simple chatbots.

It also strengthens Mistral's position as one of the most important European AI companies trying to compete with larger U.S. and Chinese model providers while maintaining a strong open-weight strategy.

What This Means for Developers in the Philippines

There is no special Philippine rollout required because the public-preview API can be accessed through Mistral's platform wherever the service is available.

For local developers, startups, and teams experimenting with AI automation, the main questions will be practical:

• how ML4 compares with existing open and closed models on real workloads;
• how much compute will be needed to self-host the weights once they are released;
• whether API pricing is competitive after the launch discount ends;
• how well the 1 million-token context performs with large codebases and documents; and
• whether the model's agent and multimodal capabilities are reliable enough for production use.

The upcoming open-weight release will be especially important for organizations that want more control over deployment and data handling.

The Bottom Line

Mistral Large 4 is now in public preview with 1.05 trillion total parameters, 49 billion active parameters, multimodal input, and a 1 million-token context window.

Mistral is positioning it as a flagship model for coding, agentic workflows, enterprise tasks, and multimodal reasoning.

The API is available now, while the company says downloadable weights will arrive by the end of October.

The bigger test will come once developers can run and evaluate those weights directly. Until then, ML4 is one of the most notable open-weight AI launches of the month, but it is still a preview rather than a finalized general-availability release.