Devstral 2

Learning Objectives

• Understand what Devstral 2 is and how it differs from general-purpose LLMs and its predecessor Codestral
• Identify Devstral 2's core differentiators: SWE-bench performance, cost efficiency, open weights, and the Mistral Vibe CLI
• Evaluate when Devstral 2 is the right choice versus GitHub Copilot, Claude for coding, or other open-weight coding models

What Is Devstral 2?

Devstral 2 is Mistral AI's dedicated software engineering model — purpose-built for real-world software development tasks rather than adapted from a general-purpose language model. Released in December 2025 as the successor to Codestral, Devstral 2 represents a shift from code completion to full software engineering — understanding codebases, planning changes across multiple files, debugging complex issues, and executing end-to-end development workflows.

Devstral 2 comes in two sizes: the full Devstral 2 (123 billion parameters) and Devstral Small 2 (24 billion parameters). The full model achieves 72.2% on SWE-bench Verified — a benchmark that tests whether a model can actually resolve real GitHub issues from popular open-source repositories — while Devstral Small 2 scores 68.0%, making it one of the most capable small coding models available.

A key claim from Mistral is that Devstral 2 is up to 7x more cost-efficient than Claude Sonnet on real-world software engineering tasks. This efficiency comes from the model's specialized training: rather than spending parameters on general conversation, creative writing, and world knowledge, Devstral 2 focuses its capacity entirely on understanding and generating code.

Devstral 2 launched alongside Mistral Vibe — a terminal-native coding agent tool that lets developers use Devstral 2 directly from the command line, without needing an IDE extension or web interface.

Pricing & Access

Devstral Small 2's Apache 2.0 license is notably more permissive than the modified MIT license on the full model — making it the preferred choice for startups and projects that need unrestricted commercial use.

Core Capabilities

Real-World Software Engineering

Unlike code completion models that suggest the next line, Devstral 2 handles complete software engineering workflows:

• Issue resolution: Given a GitHub issue or bug report, analyze the codebase, identify the root cause, and generate a multi-file fix — this is what SWE-bench measures
• Multi-file changes: Plan and execute changes that span multiple files, understanding how modifications in one file affect imports, tests, and dependent code elsewhere
• Codebase understanding: Navigate large repositories, understand architectural patterns, and make changes that respect existing conventions and patterns

The 72.2% SWE-bench Verified score means Devstral 2 can resolve nearly three-quarters of real-world GitHub issues autonomously — a significant improvement over Codestral's more limited code completion capabilities.

256K Context Window

With 256K tokens of context, Devstral 2 can ingest substantial portions of a codebase:

• Repository-scale understanding: Load dozens of files simultaneously to understand cross-file dependencies
• Long file analysis: Process large configuration files, data schemas, or monolithic modules without truncation
• Extended agent sessions: Multi-step coding agents can maintain context across many planning and execution cycles

Mistral Vibe CLI

Launched alongside Devstral 2, Mistral Vibe is a terminal-native coding agent:

• No IDE required: Run coding tasks from any terminal — plan features, write code, fix bugs, and run tests
• Agent workflow: Vibe operates as an autonomous agent — it reads your codebase, plans changes, writes code, and can run commands to verify its work
• Developer-first: Designed for developers who prefer terminal workflows over GUI-based tools

Cost Efficiency

Devstral 2's specialized focus translates directly to cost savings:

• Up to 7x cheaper: Mistral claims up to 7x cost efficiency over Claude Sonnet on real-world software engineering tasks
• Smaller model, focused capability: At 123 billion parameters (or 24 billion for Small), Devstral 2 runs on less compute than frontier general-purpose models while matching or exceeding their coding performance
• Token efficiency: Specialized training means fewer wasted tokens on reasoning about non-code knowledge

Strengths

• Top-tier SWE-bench: 72.2% on SWE-bench Verified (full model) — resolves real GitHub issues, not just generates code snippets
• Cost-efficient: Up to 7x cheaper than comparable models on real-world coding tasks
• Two model sizes: 123 billion for maximum capability, 24 billion for lightweight deployment — both open weights
• Apache 2.0 small model: Devstral Small 2 is fully permissively licensed — no commercial use restrictions
• Terminal-native agent: Mistral Vibe CLI enables coding agent workflows directly from the command line
• 256K context: Handles large codebases and multi-file changes across extended agent sessions
• European data handling: Code processed through Mistral's EU infrastructure for teams with data sovereignty requirements

Limitations & Considerations

• Coding-focused only: Not designed for general conversation, analysis, or creative writing — use Mistral Large 3 for non-coding tasks
• Less ecosystem than Copilot: GitHub Copilot has deeper GitHub integration (PR summaries, issue references, Copilot Chat in github.com) that Devstral 2 does not replicate
• Modified MIT on full model: The 123 billion model's modified MIT license requires review for some commercial use cases — the 24 billion Apache 2.0 model avoids this
• Self-hosting compute: The full 123 billion model requires multi-GPU infrastructure; the 24 billion model is more accessible but trades some capability
• Newer model: As a December 2025 release, community tooling and integrations are still maturing compared to GitHub Copilot's established ecosystem

Best Use Cases

When to choose alternatives:

• Want the deepest GitHub integration → GitHub Copilot (PR summaries, issue linking, github.com chat)
• Need code + general reasoning in one model → Claude Opus or GPT-5.1
• Want fully managed subscription pricing → GitHub Copilot ($10-39/month, unlimited completions)
• Broader language model with coding capability → Mistral Large 3 for combined coding and general tasks

Getting Started

1. Sign up for Mistral — create an account at console.mistral.ai and generate an API key
2. Install Mistral Vibe — follow the installation guide at docs.mistral.ai for npm or pip
3. Run your first task — navigate to a project directory and ask Vibe to explain the codebase, fix a bug, or add a feature
4. Try IDE integration — install Continue.dev in VS Code or JetBrains and select Devstral 2 as your model
5. Test SWE-bench-style tasks — give Devstral 2 a real bug report or feature request from your project and evaluate its multi-file solution
6. Compare models — test the same coding tasks with GitHub Copilot or Claude to determine which fits your workflow

Key Takeaways

• Devstral 2 is Mistral AI's dedicated software engineering model — superseding Codestral with a focus on real-world issue resolution (72.2% SWE-bench Verified) rather than just code completion
• Available in two sizes (123 billion modified MIT, 24 billion Apache 2.0) with up to 7x cost efficiency over comparable models — making frontier coding capability accessible at lower cost
• Launched alongside Mistral Vibe CLI, a terminal-native coding agent that brings autonomous software engineering to the command line
• Best for autonomous bug fixing, cost-sensitive coding workflows, and self-hosted deployment; choose GitHub Copilot for ecosystem integration or Claude/GPT for combined coding and general reasoning