Repository-scale code changes
Explore a codebase, coordinate edits across files, run tests, and correct failures.
Independent tool overview
Devstral 2 at a glance
Devstral 2 is Mistral's open-weight coding-model family for repository exploration, multi-file editing, and software-engineering agents, offered as a 123B model and a more locally deployable 24B Small variant.
Visit the official Devstral 2 site ↗%2520(1).png&w=3840&q=75)
- Model family
- Devstral 2 123B and Devstral Small 2 24B
- Context window
- 256K tokens for both variants
- Primary use
- Agentic software engineering
- Licenses
- Modified MIT for 123B; Apache 2.0 for Small 2
- Official harness
- Mistral Vibe CLI
- Hosted API
- Original endpoint deprecated; Medium 3.5 is the replacement
Overview
What Devstral 2 is
Devstral 2 is a specialized coding-model family released by Mistral AI with All Hands AI. The 123B dense model targets autonomous software engineering in data-center environments, while Devstral Small 2 brings related agentic coding capabilities to a 24B checkpoint that can run on far smaller hardware.
Both variants support a 256K context and are designed to use tools to search repositories, edit multiple files, execute commands, run tests, detect failures, and retry. Mistral's open-source Vibe CLI supplies a native terminal harness with project context, file and shell tools, subagents, permission controls, and IDE integration.
The model weights remain available, but the original hosted Devstral 2 API version is deprecated for new integrations as of May 22, 2026. Mistral directs new hosted coding integrations to Mistral Medium 3.5, which also powers the newer remote-agent experience in Vibe.
Use cases
Who Devstral 2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Private coding agents
Keep source code within controlled infrastructure by serving open weights locally or on-premises.
Legacy modernization
Analyze architecture and incrementally migrate older code with test-driven verification.
Custom coding workflows
Pair the weights with Vibe, Cline, Kilo Code, OpenHands, or another compatible agent scaffold.
Local experimentation
Use the 24B Small variant when the 123B model's data-center requirements are impractical.
Capabilities
Core Devstral 2 features
Agentic code editing
Built to inspect repositories, modify multiple files, and use development tools across a task.
256K context
Handles large code and conversation histories, with input and output sharing the total limit.
Function calling
Supports tool-driven workflows and structured software-engineering agents.
Two model sizes
Offers a stronger 123B data-center model and a smaller 24B checkpoint for local deployment.
Open weights
Publishes FP8 weights for both family members with model-specific licenses.
Mistral Vibe CLI
Provides project-aware terminal automation, file references, commands, history, subagents, and configurable permissions.
Private deployment
Supports on-premises serving and enterprise customization without sending code to a shared hosted model.
Small-model vision input
Devstral Small 2 accepts images for multimodal coding agents; the 123B model is text-to-text.
Process
How the Devstral 2 workflow works
Step 1
Choose the current path
Use open weights for Devstral 2 itself, or Mistral Medium 3.5 for a new Mistral-hosted integration.
Step 2
Match model to hardware
Reserve the 123B model for data-center GPUs and use Small 2 when single-system deployment is the goal.
Step 3
Select an agent scaffold
Configure Vibe or another compatible harness with repository, terminal, Git, and testing tools.
Step 4
Set strict permissions
Require approval for destructive commands, network access, secrets, deployment, and irreversible changes.
Step 5
Start with a bounded issue
Give acceptance criteria, relevant tests, repository guidance, and explicit files or systems that are out of scope.
Step 6
Inspect the complete diff
Review architecture, dependencies, security, tests, generated artifacts, and unrelated modifications.
Step 7
Benchmark on real work
Compare pass rate, review time, regressions, latency, compute, and total trajectory cost before scaling.
Cost
Devstral 2 pricing and free plan
The original Devstral 2 hosted endpoint is deprecated, so its launch price is no longer the right basis for a new API integration. Open weights remain downloadable, with infrastructure and license obligations. Mistral's current hosted replacement is Medium 3.5.
Devstral 2 open weights
Free weights under Modified MIT
Self-host the 123B FP8 checkpoint for private coding-agent workloads.
- Modified MIT revenue restriction applies
- Companies above $20M consolidated monthly revenue need separate commercial permission
- Mistral recommends at least four H100-class GPUs
- Infrastructure and operations are separate
Devstral Small 2 open weights
Free under Apache 2.0
Use the 24B checkpoint for more accessible local and private deployment.
- Single-GPU deployment is supported
- CPU-only configurations are possible with slower performance
- Supports image input
- Serving and electricity costs apply
Original Devstral 2 API
Deprecated for new integrations
The hosted 25.12 endpoint was priced at $0.40 input and $2 output per million tokens before deprecation.
- Deprecated May 22, 2026
- Do not start a new dependency on the old model ID
- Free experimentation was previously offered through Mistral Studio
Mistral Medium 3.5 replacement
$1.50 input / $0.15 cached input / $7.50 output per 1M tokens
Mistral's current hosted replacement for new agentic coding integrations.
- 256K context
- Multimodal input
- Powers current Vibe remote agents
- Batch and regional options can change effective cost
Pricing checked . Check current pricing at the source ↗
Assessment
Devstral 2 strengths and limitations
Where it stands out
- Specialized for repository exploration, multi-file edits, tool use, and test-driven correction
- Two model sizes create practical choices between capability and local deployment
- Small 2's Apache 2.0 license is straightforward for commercial self-hosting
- Open weights enable private source-code workflows and custom serving
- The official Vibe CLI provides a capable, configurable agent harness
- 256K context supports substantial repository and task history
- Mistral publishes the current replacement path instead of leaving new API users on a stale endpoint
What to consider
- The original hosted Devstral 2 API version is deprecated for new integrations
- The 123B model's Modified MIT license excludes organizations above $20M consolidated monthly revenue without separate permission
- Mistral recommends at least four H100-class GPUs for the 123B checkpoint
- Devstral Small 2 is easier to run but does not match the larger model's full capability
- The 123B variant is text-only; only Small 2 adds image input
- A 256K limit includes both prompt and generated output and does not guarantee perfect repository recall
- Agentic coding still exposes the repository and execution environment to destructive or insecure actions
- Mistral's own launch evaluation acknowledged that a capability gap with leading closed models remained
Compare
Devstral 2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
MiMo-V2.5-Pro
Choose MiMo-V2.5-Pro for a newer MIT-licensed coding model with a one-million-token context and low hosted token prices.
Explore MiMo-V2.5-Pro →Coding
Kimi K2.7 Code
Evaluate Kimi K2.7 Code for another current open-weight model built specifically for coding-agent trajectories.
Explore Kimi K2.7 Code →Business Operations
DeepSeek
Consider DeepSeek for a broader low-cost model family with coding, reasoning, hosted API, and open-weight options.
Explore DeepSeek →Consumer
Trinity-Large-Thinking
Use Trinity Large Thinking for permissive long-context agentic work beyond a coding-only specialization.
Explore Trinity-Large-Thinking →Questions
Devstral 2 FAQs
What is Devstral 2?
It is Mistral's coding-model family for agents that explore repositories, edit multiple files, execute tools, run tests, and correct failures.
Is Devstral 2 still active?
The open weights remain available and the model family remains active, but the original hosted Devstral 2 API version is deprecated for new integrations. Mistral recommends Medium 3.5 for new hosted work.
What is the difference between Devstral 2 and Small 2?
Devstral 2 is a 123B text model for data-center deployment. Small 2 is a 24B model that can run on much smaller hardware and also supports image input.
What licenses apply?
The 123B model uses a Modified MIT license with a $20M consolidated monthly-revenue restriction. Small 2 uses Apache 2.0. Vibe CLI is also Apache 2.0.
How much hardware does Devstral 2 need?
Mistral recommends a minimum of four H100-class GPUs for the 123B model. Small 2 is designed for single-GPU and even CPU-only configurations.
What is the context window?
Both Devstral 2 variants support 256K total context, shared between input and generated output.
What replaced the hosted Devstral 2 API?
Mistral's model documentation directs new integrations to Mistral Medium 3.5, a 256K multimodal agentic and coding model.
What is Mistral Vibe?
Vibe is Mistral's open-source coding agent for the terminal and IDE, with repository context, development tools, subagents, skills, permission controls, and remote-agent options.
Bottom line
Our Devstral 2 verdict
Devstral 2 remains useful as an open-weight coding family, especially when source code must stay private. Small 2 is the practical choice for accessible self-hosting, while the 123B checkpoint demands expensive infrastructure and careful license review. For a new hosted integration, follow Mistral's current guidance and evaluate Medium 3.5 instead of building on the deprecated Devstral endpoint. In every path, the agent harness, permissions, tests, and diff review matter as much as the model.
Visit Devstral 2 website ↗Similar Coding Tools
View all Coding AI tools →
Mistral Vibe CLI
Mistral Vibe CLI - Mistral's new open-source CLI coding assistant powered by its Devstral models
Devin Desktop (formerly Windsurf)
Codeium Windsurf: Provides ai-powered code suggestions, completions, and debugging tools for developers.
Codex Max
Codex Max - OpenAI's new frontier agentic coding model
GPT-5.2-Codex
GPT-5.2-Codex - Agentic coding model for professional software engineering and defensive cybersecurity.
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.