Devstral 2

From Mistral AI

Mistral AI’s State of the Art Open-Source Agentic Coding Model Leading Performance in Software Engineering Tasks

Pricing Model

Starting Price

$14.99/Month

Visit Devstral 2
Last Updated: December 15, 2025
By Zelili AI

About This AI

Devstral 2 is Mistral AI’s next-generation coding model family released on December 9, 2025, available in Devstral 2 (123B parameters) and Devstral Small 2 (24B parameters).

It is a dense transformer optimized for agentic software engineering, supporting codebase exploration, multi-file edits, architecture-level context, and tool orchestration for autonomous code agents.

With a 256K token context window, it enables processing large repositories in one pass while maintaining coherence.

The model achieves 72.2% on SWE-bench Verified (SOTA among open-weight models), outperforming many peers in real-world GitHub issue resolution and software reasoning benchmarks.

Released under modified MIT license (Devstral 2) and Apache 2.0 (Small variant), it is fully open-source with weights on Hugging Face for local deployment.

Currently free via Mistral API during initial period; post-free pricing is $0.40/$2.00 per million tokens (input/output) for the 123B version and lower for Small.

Paired with Mistral Vibe CLI (terminal-native agent), it enables end-to-end code automation, custom subagents, slash-command skills, and workflow configuration.

Ideal for developers, AI agents, software teams, and researchers building autonomous coding systems, with strong cost-efficiency (up to 7x better than some closed models) and local runnability on consumer hardware for the Small version.

Key Features

  1. 256K token context window: Handles entire large codebases and long conversations without truncation
  2. Agentic coding excellence: Explores repositories, edits multiple files, maintains architecture context, and orchestrates changes
  3. SOTA open-weight performance: 72.2% on SWE-bench Verified for real-world software engineering tasks
  4. Multi-size variants: 123B dense for maximum capability; 24B Small for local/consumer hardware deployment
  5. Tool use and reasoning: Built for autonomous agents with reliable tool calling and multi-step planning
  6. Open-source availability: Full weights under modified MIT/Apache 2.0, easy Hugging Face integration
  7. Mistral Vibe CLI integration: Native terminal agent with custom subagents, skills, and workflows
  8. High cost-efficiency: Up to 7x more efficient than competitors like Claude Sonnet in real tasks
  9. Strong benchmarks: Competitive or leading in coding, reasoning, and agentic evaluations vs open peers

Price Plans

  1. Free (Current API period $0): Free access via Mistral API during launch period; also free on Experiment plan in Mistral Studio
  2. API Paid (Post-free $0.40/$2 per M tokens input/output for 123B): Standard token pricing after promo; lower for Small 2 ($0.10/$0.30)
  3. Le Chat Pro/Team ($14.99–$24.99/Month): Access to Devstral 2 via Vibe CLI, higher limits, and integrated coding workflows

Pros

  1. Top open coding model: Sets new SOTA for open-weight agentic coding with high SWE-bench score
  2. Long context support: 256K tokens enable full-repo understanding without splitting
  3. Free initial access: Currently free via API; Small variant runs locally on consumer GPUs
  4. Strong efficiency: Delivers high performance at lower cost than many closed models
  5. Developer-focused: Pairs perfectly with Mistral Vibe for terminal-native automation
  6. Open weights: Full customization, fine-tuning, and local deployment possible
  7. Rapid innovation: Builds on Mistral's fast release cadence for coding AI

Cons

  1. High compute for full model: 123B requires significant GPUs (e.g., 4+ H100 equivalents)
  2. Free period temporary: API moves to paid ($0.40/$2 input/output per M tokens) after initial offer
  3. Verbose outputs: Can produce longer responses unless prompted for conciseness
  4. Setup for local use: Small variant easier, but full needs enterprise hardware
  5. No native multimodal: Primarily text/code-focused; limited vision compared to some rivals
  6. Recent release: Adoption metrics still emerging; no massive user base numbers yet

Use Cases

  1. Autonomous code agents: Build agents that explore repos, fix bugs, refactor code across files
  2. Software engineering tasks: Resolve GitHub issues, implement features, review PRs
  3. Local development: Run Small variant on laptop for offline coding assistance
  4. CLI automation: Use Mistral Vibe for terminal-based end-to-end workflows
  5. Research in agentic AI: Fine-tune or extend for custom software agents
  6. Team productivity: Integrate into IDEs or CI/CD for faster iteration
  7. Cost-effective coding: High performance at fraction of closed-model costs

Target Audience

  1. Software developers: Needing agentic help for coding, debugging, and refactoring
  2. AI agent builders: Creating autonomous software engineering systems
  3. Open-source enthusiasts: Running and customizing top open coding models locally
  4. Dev teams: Boosting productivity with integrated CLI agents and API
  5. Researchers: Benchmarking or advancing agentic coding frontiers
  6. Startups/enterprises: Seeking efficient, open alternatives to proprietary LLMs

How To Use

  1. Access via API: Sign up at console.mistral.ai, get API key, and call Devstral 2 endpoint
  2. Try in Mistral Studio: Use free Experiment plan or Le Chat Pro for immediate testing
  3. Run locally (Small): Download weights from Hugging Face, set up with transformers/vLLM
  4. Use Mistral Vibe CLI: Install via pip/npm, configure with API key or local model
  5. Prompt for coding: Ask agentic tasks like 'fix this bug in repo X' or 'refactor module Y'
  6. Integrate in tools: Use in IDEs via extensions or CLI for seamless workflow
  7. Monitor usage: Track tokens in console; switch to Small for cost/latency savings

How we rated Devstral 2

  • Performance: 4.8/5
  • Accuracy: 4.7/5
  • Features: 4.8/5
  • Cost-Efficiency: 4.9/5
  • Ease of Use: 4.5/5
  • Customization: 4.9/5
  • Data Privacy: 4.8/5
  • Support: 4.4/5
  • Integration: 4.7/5
  • Overall Score: 4.7/5

Devstral 2 integration with other tools

  1. Mistral API: Direct integration for web/apps via console.mistral.ai endpoints
  2. Mistral Vibe CLI: Native terminal agent for local or API-powered end-to-end coding automation
  3. Hugging Face: Model weights and inference pipelines for local deployment and fine-tuning
  4. IDE Extensions: Compatible with VS Code, Cursor, Zed via CLI or API wrappers
  5. Agent Frameworks: Works with LangChain, LlamaIndex, AutoGen for custom agentic systems

Best prompts optimised for Devstral 2

  1. Explore this GitHub repository [repo URL], understand the architecture, and suggest improvements to the main module while maintaining compatibility
  2. Fix the failing test in file tests/unit.py by editing the function in src/core.py; reason step-by-step and apply changes
  3. Refactor this legacy codebase to use modern async patterns in Python; update multiple files and add type hints
  4. Implement a new feature: add user authentication with JWT to this Express.js app; include routes, middleware, and tests
  5. Analyze this pull request diff and provide a detailed code review with suggestions for security and performance
Devstral 2 sets a new benchmark for open-source agentic coding models with 72.2% on SWE-bench Verified and a massive 256K context window. The 123B version delivers frontier performance, while Small 2 enables local use. Free initial API access and strong cost-efficiency make it ideal for developers building autonomous agents and software tools.

FAQs

  • What is Devstral 2?

    Devstral 2 is Mistral AI’s next-generation open-source coding model family (123B and 24B variants) optimized for agentic software engineering tasks, released December 9, 2025.

  • When was Devstral 2 released?

    It was officially released on December 9, 2025, alongside Mistral Vibe CLI for terminal-native coding automation.

  • What is the performance of Devstral 2?

    Devstral 2 achieves 72.2% on SWE-bench Verified, setting SOTA among open-weight models for real-world software engineering benchmarks.

  • Is Devstral 2 free to use?

    Currently free via Mistral API during initial period; post-free API pricing is $0.40/$2 per million tokens (input/output) for the 123B version.

  • What is the context window for Devstral 2?

    It supports a 256,000 token context window, enabling full codebase analysis in a single pass.

  • How does Devstral 2 compare to other models?

    It outperforms many open peers on SWE-bench and is more cost-efficient than closed models like Claude Sonnet for real tasks.

  • Can Devstral Small 2 run locally?

    Yes, the 24B Devstral Small 2 is designed for consumer hardware deployment with the same long context and strong performance.

  • What is Mistral Vibe?

    Mistral Vibe is a native CLI agent powered by Devstral 2 for end-to-end code automation, custom workflows, and terminal integration.

Newly Added Tools​

Qwen-Image-2.0

$0/Month

Qodo AI

$0/Month

Codiga

$10/Month

Tabnine

$59/Month
Devstral 2 Alternatives

Qodo AI

$0/Month

Codiga

$10/Month

Tabnine

$59/Month

About Author

Hi Guys! We are a group of ML Engineers by profession with years of experience exploring and building AI tools, LLMs, and generative technologies. We analyze new tools not just as a user, but as someone who understands their technical depth and real-world value.We know how overwhelming these tools can be for most people, that’s why we break down complex AI concepts into simple, practical insights. Our goal is to help you discover these magical AI tools that actually save your time and make everyday work smarter, not harder.“We don’t just write about AI: We build, test and simplify it for you.”