Mistral Large 4, also known as "Le Chonk", uses 1 trillion total parameters with 49 billion active parameters. This model supports a 1-million token context window and up to 256K output tokens. Mistral provides two methods for building agents on La Plateforme. The Agent Builder provides a user interface for model selection, sampling temperature adjustments, and adding instructions or few-shot learning examples. The Agent API provides a programmatic solution for developers to integrate agent creation into existing workflows. Developers choose from various models, such as Mistral Large 2 (mistral-large-2407), Mistral Nemo (open-mistral-nemo), or Codestral (codestral-2405). Mistral Large 4 delivers results for workloads in cyber defense, manufacturing, and finance. The model is available via API immediately, while the open weights release occurs at the end of October. Once a user deploys an agent, they interact with it via API using an agent_id or enable chat functionality on Le Chat. This flexibility allows for communication and collaboration within business operations.
Enterprise customization and regional control
Mistral Forge allows enterprises and governments to train models from the ground up using internal data. This approach helps companies avoid reliance on third-party providers and handle domain-specific or non-English data. Forge uses a library of open-weight models, including the Mistral Small 4. Mistral engineers collaborate with clients to adapt data and implement models, and they provide tools for synthetic data pipelines. Mistral has already partnered with organizations including Ericsson, the European Space Agency, and the consulting firm Reply. Early adopters include the Dutch chipmaker ASML. Mistral also provides regional endpoints so customers decide if inference runs in Europe or the United States. This helps organizations manage data residency, regulatory requirements, and latency. You already know that data sovereignty matters for compliance, so these regional controls provide a way to keep sensitive data in specific jurisdictions.
Mistral provides the Vibe interface for various environments. The Vibe Web interface allows users to execute tasks from a chat interface. The CLI Vibe Agent is a native terminal agent for engineering tasks. The Vibe IDE extension works for VS Code and JetBrains. The Vibe mobile app for iOS and Android includes voice mode, web research, document libraries, and task planning. Users can also use Vibe to create, edit, and share spreadsheets or presentations. The mobile application handles up to 30 GB of storage per user for Enterprise customers.
| Feature | Free Plan | Pro Plan | Team Plan | Enterprise Plan |
|---|---|---|---|---|
| Web Research | Up to 5x | Up to 5x | Unlimited | Unlimited |
| Task Planning | Not listed | Not listed | Unlimited | Unlimited |
| Storage | Not listed | Not listed | Up to 15 GB | Up to 30 GB |
| Project Folders | Not listed | Not listed | Up to 1,000 | Up to 1,000 |
The Team plan costs $24.99 per user per month. Mistral Large API costs $0.50 per million incoming tokens and $1.50 per million outgoing tokens. The Priority Tier, in public preview, costs 1.75 times the standard list price and provides a 99.5% uptime SLA.
Managed workflows and deployment limits
Managed Deployments, in public preview, runs Workflows on Mistral Cloud infrastructure in the nl-north-1 region. This service requires a paid plan and uses service-account authentication. Pay-as-you-go users can run three concurrent deployments, while Enterprise users can run ten. Users build Docker images from GitHub repositories and run them according to an ENTRYPOINT or CMD. The platform provides OpenTelemetry logs, traces, and metrics to monitor performance. The platform provides OpenTelemetry logs, traces, and metrics that allow engineers to examine the timing and errors of every execution to ensure the workflow completes as expected throughout the entire deployment process.
The service also provides an automatically rotating service-account token to the worker. This token is used for authentication in the managed deployment. Developers must manage secrets carefully because changing a secret value does not restart the worker. Sensitive values in Docker build arguments can remain in image history and layers. This creates a security risk for teams using automated builds. The service requires Workflows SDK 3.10 or later for support.
How do teams effectively manage the lifecycle of secrets when workers do not automatically restart after a value change?




