Mistral Coding Stack: DevStral 2, Vibe, and Medium 3.5 for Enterprise Teams

Mistral has rebuilt its enterprise coding offering around DevStral 2, Vibe (formerly Le Chat), and Medium 3.5, consolidating models, agents, and IDE surfaces into a single stack. Designed to address compliance and security concerns, this setup offers self-hosted execution, EU data residency, and signed audit logs for regulated engineering teams.

What is DevStral 2?

DevStral 2 is Mistral's agentic coding model family, developed in partnership with All Hands AI. It is available in two main configurations:

• DevStral 2 (123B dense): Featuring a 256K context window for production enterprise applications.
• DevStral Small 2 (24B): Retaining the 256K context and agentic tuning, designed to run on a single GPU or a local development machine with 32GB RAM.
• Both models are released under the Apache 2.0 license and can be accessed via Hugging Face, Ollama, and the Mistral API.

In terms of performance benchmarks, DevStral 2 builds on DevStral Small's 46.8% score on SWE-Bench Verified. API pricing is set at $0.40 per million input tokens and $2.00 per million output tokens for the flagship model, while DevStral Small 2 costs $0.10 per million input and $0.30 per million output tokens.

The Mistral Coding Stack

The overall coding suite is packaged as a unified enterprise offering:

• DevStral 2 and DevStral Small 2: Main engines for repository-scale code generation and agentic tasks.
• Codestral: Optimized for fast, single-file completions and fill-in-the-middle operations.
• Mistral Medium 3.5: A generalist 128B open-weight model scoring 77.6% on SWE-Bench Verified. It runs on four GPUs and handles both reasoning and coding within a single set of weights.
• Vibe: The user-facing agent interface, formerly known as Le Chat.

For procurement, Mistral bundles Codestral, DevStral 2, Vibe, and Le Chat Enterprise with EU hosting, self-hosted deployment options, and the Forge platform for custom fine-tuning.

Vibe 2.0: The Coding Agent

Mistral has positioned the Vibe product line into two separate tools: Vibe for Work (the standard chat interface, formerly Le Chat) and Vibe for Code (the agent).

Vibe 2.0 functions as a terminal-native coding agent powered by DevStral 2, incorporating several key features:

• Custom Subagents: Designed for specific workflows like PR reviews and test generation.
• Multi-Choice Clarifications: Prompts the user with explicit choices when intent is ambiguous instead of making assumptions.
• Slash-Command Skills: Loads preconfigured workflows for deployment, linting, or database migrations.
• Unified Agent Modes: Groups tools, permissions, and behaviors into named profiles.
• Continuous Updates: Delivers automatic, non-disruptive bug fixes.

The command-line interface, VS Code extension, JetBrains plugin, and Zed plugin share a unified harness, allowing users to start a session in the terminal and transition directly into their IDE without losing context.

Mistral also offers Vibe remote agents, which provide cloud sandboxes to run asynchronous coding tasks in parallel. Local CLI sessions can be moved to the cloud via the /teleport command, with integrations available for GitHub, Linear, Jira, Sentry, Slack, and Microsoft Teams. Remote agents run on Mistral Medium 3.5 and cost $1.50 per million input tokens and $7.50 per million output tokens.

Enterprise Safeguards and Governance

The suite includes specific compliance and security features designed for enterprise environments:

• EU Data Residency: Mistral Cloud hosts data in the EU by default. Self-hosted deployments allow teams to keep weights and prompts entirely inside their own VPC.
• Role-Based Policies: Control which developer groups are authorized to run destructive commands, open pull requests, or trigger production deployments.
• Audit Logging: Generates signed logs alongside human-readable summaries to help security teams audit code changes.
• Network and Secret Controls: Restricts the external hosts the agent can access and routes credentials directly through existing secrets managers.
• Model Provenance: Records the model name, version, and temperature settings for every run to ensure reproducibility.
• Open Weights: DevStral 2's Apache 2.0 license and Medium 3.5's modified MIT license prevent vendor lock-in.

Getting Started: Running DevStral 2

To test the system in a sandboxed environment:

1. Install the CLI: Run the official installation script or use uv tool install and connect your API key.
2. Define a Mode: Create an agent mode pointing to a non-production branch, limiting permissions to formatting, testing, and documentation.
3. Map the Project: Allow the agent to scan local files (like package configurations, API specifications, or Infrastructure-as-Code layouts) to build a project context map.
4. Execute a Sandbox Task: Run a targeted task, such as refactoring a piece of middleware or generating test templates, and review the proposed execution plan beforehand.
5. Review Output: Inspect plans, diffs, and logs in your monitoring tools before promoting workflows to CI using a scheduled remote agent.

Who Is This Stack For?

The DevStral 2 and Vibe stack is best suited for organizations balancing compliance, cost, and developer autonomy:

• Enterprise Engineering: Keeps critical repositories secure while providing developers with an agent that respects branch protection policies.
• Platform Engineering: Standardizes how AI agents interact with local build systems, observability stacks, and credential managers.
• Data and ML Teams: Automates data pipeline reviews, documentation, and experimental scripting using DevStral 2's code-focused reasoning.
• Security Teams: Standardizes development within a controlled framework using EU-hosted endpoints or local deployments rather than unmonitored tools.
• Highly Regulated Industries: Provides banking, healthcare, and public sector organizations with GDPR-compliant defaults out of the box.

Comparative Overview

The following table compares the Mistral coding stack against other prominent enterprise options:

Platform	Governance	Runtime	Model Bias	Best Fit
DevStral 2 + Vibe	Policy modes, signed logs, EU hosting, self-hosting	CLI, VS Code, JetBrains, Zed	DevStral 2 + Medium 3.5	Regulated teams requiring self-hosted control
Google AntiGravity	Mission control artifacts	VS Code fork with browser integration	Gemini + Claude	Teams prioritizing multi-agent UI workflows
OpenAI Codex	Cloud sandboxes, organization policies	Cloud + ChatGPT-tied CLI	GPT-5 family	OpenAI-native organizations using ChatGPT Enterprise

For organizations constrained by strict compliance or data sovereignty requirements, DevStral 2's open weights, self-hosted deployment models, and cross-IDE parity represent its primary differentiators.

FAQ

What is DevStral 2?
DevStral 2 is Mistral's agentic coding model family, built in collaboration with All Hands AI. It includes a 123B dense model with a 256K context window for production use and a 24B Small version designed for local development.

Who should adopt the Mistral coding stack?
The stack is designed for enterprise development teams, platform engineers, and security groups that require EU data residency, self-hosted weights, and transparent, auditable code generation processes.

What changed in Vibe 2.0?
Vibe 2.0 updates the assistant interface into a terminal-native coding agent. It introduces custom subagents, slash-command skills, multi-choice intent clarification, unified permissions, and cross-plugin synchronization.

How does Mistral compare to other enterprise AI coding platforms?
Mistral focuses on open weights, EU hosting, self-hosted VPC deployments, and parity between CLI and IDE environments. In contrast, Google AntiGravity provides a multi-agent UI built into a VS Code fork, and OpenAI Codex relies on cloud sandboxes connected to ChatGPT accounts.

Summary

Mistral's coding stack provides the core components that enterprise procurement and security teams look for: open-weight options, self-hosted deployment, strict policy controls, and a developer-friendly interface that integrates into existing terminal and IDE setups. For teams in highly regulated sectors where data residency has previously blocked the adoption of AI coding assistants, DevStral 2 and Vibe 2.0 offer a viable, self-managed alternative.