Devstral 2

Devstral 2 Overview

Mistral released Devstral 2 in December 2025 as a dense transformer specialized for code agents. The model can supply reasoning and tool decisions, but the agent harness supplies repository access, file editing, version-control operations, command execution, and the feedback loop.

Devstral 2 is optimized for data-center GPUs. Mistral's release guidance states that self-deployment requires a minimum of four H100-class GPUs. The smaller Devstral Small 2 targets single-GPU and consumer-hardware configurations.

Mistral's current API pricing page, rechecked on July 17, 2026, lists Devstral 2 at $0.40 per million input tokens and $2.00 per million output tokens. Devstral Small 2 is listed at $0.10 input and $0.30 output. Hosted cost is only one part of the choice; larger-model quality may reduce retries on complex changes.

SWE-bench 72.2% Breakdown

Mistral reports 72.2% on SWE-bench Verified for Devstral 2. This is a vendor-published result for a model operating through an agent scaffold, not an isolated measure of the weights.

SWE-bench Verified tests whether a system can resolve selected real GitHub issues. Results depend on:

• repository setup and test environment;
• file and terminal tools exposed to the model;
• prompt and system instructions;
• maximum turns, token budget, and retries;
• patch application and validation logic;
• sampling settings and agent version.

Use the 72.2% figure as evidence that Mistral evaluated the model for repository-level work, not as a guarantee that 72.2% of your issues will be fixed. A local acceptance suite should include the task types, languages, frameworks, and test conventions your team actually uses.

Devstral 2 vs Devstral Small 2

Choose Devstral 2 when difficult multi-file tasks justify the larger serving footprint or hosted price. Choose Small 2 when local privacy, lower latency, lower cost, or easier customization outweighs the benchmark gap.

Run both on the same repository suite. A four-point benchmark difference can matter less than whether one model consistently follows your build tools and project conventions.

Multi-File Editing Capabilities

Multi-file editing requires the model to preserve relationships across interfaces, tests, configuration, migrations, and documentation. Mistral describes Devstral 2 as able to explore codebases, orchestrate changes across files, detect failures, and retry with corrections.

A strong evaluation task should require more than changing several files mechanically. For example:

1. modify a public interface;
2. update all callers and types;
3. add or repair tests;
4. run targeted checks and the broader suite;
5. return a focused diff with no unrelated formatting changes.

Score whether the agent finds the full dependency chain, not just whether it produces a large patch. Also record how often it edits generated files, misses configuration, or repairs one test by weakening another.

For a different xAI coding model and alias lifecycle, see the Grok Code Fast 1 guide.

License & Commercial Use

Devstral 2 uses a modified MIT license. Devstral Small 2 uses Apache 2.0. Do not transfer Small 2's license label to the full model.

The modified license is intended to be permissive, but teams should read the exact file shipped with the selected checkpoint. Commercial review should cover:

• use, modification, and distribution rights;
• notice or attribution requirements;
• any conditions added to the standard MIT text;
• third-party dependencies and serving software;
• use of model output in the target product.

This is a technical summary, not legal advice. Preserve the checkpoint revision and license text in the deployment record.

Using Devstral 2 in Verdent

Devstral 2 does not appear in Verdent's built-in lineup. An OpenRouter account may expose the endpoint through BYOK, but availability depends on that provider catalog and is not native Mistral support from Verdent.

A practical evaluation is to assign Devstral 2 one multi-file Worker task while a current built-in model reviews the final diff. Keep the repository commit, task prompt, tool permissions, and tests fixed. Compare the full model with Small 2 or another coding option on accepted-task cost rather than token price alone.

After a key and endpoint validate, Verdent can use the model in chat, presets, and delegated agent roles. Smart Suggestions and automatic compression remain built-in-model features rather than BYOK features.

Sources: Review the Devstral 2 release and current Mistral API pricing.