← Models

Devstral 2

Mistral AI · Devstral

Mistral's open-weight coding model built for agentic software work and terminal-first execution.

Type
language
Context
256K tokens
Max Output
256K tokens
Status
current
Input
$0.4/1M tok
Output
$2/1M tok
API Access
Yes
License
Apache 2.0
coding agentic open-weights fill-in-the-middle function-calling long-context
Released December 2025 · Updated April 13, 2026

Overview

Freshness note: Model capabilities, limits, and pricing can change quickly. This profile is a point-in-time snapshot last verified on April 13, 2026.

Devstral 2 is Mistral AI’s open-weight coding model for agentic software work. It sits in the part of the lineup aimed at developers who want strong code-generation and code-editing behavior, but do not want to give up model portability, self-hosting options, or lower-level control over the stack.

The release matters beyond the model alone because Mistral launched it alongside Mistral Vibe, making it clear that this is meant to power end-to-end coding workflows rather than just benchmark demos.

Capabilities

Devstral 2 is positioned for real software tasks: repository exploration, bug fixing, multi-file implementation, structured tool use, and edit-heavy coding flows. Official Mistral docs list support across chat completions, function calling, structured outputs, predicted outputs, agents and conversations, built-in tools, prefix, batching, and fill-in-the-middle generation.

That combination is unusually relevant for coding. Fill-in-the-middle matters for editor-style completion and patching, while structured outputs and function calling matter for agent frameworks and coding CLIs that need reliable tool invocation rather than freeform prose.

Technical Details

Mistral’s official Devstral 2 page publishes a 256K context window and prices the model as an open release under the 25.12 generation. The docs do not foreground a separate completion ceiling in the same way some providers do, so this profile records maxOutput as the same 256K snapshot limit for repo comparability rather than as a distinct provider-published cap.

The official feature matrix also shows broad product-surface compatibility instead of a narrow “completion only” release, which is a strong signal that Mistral expects Devstral 2 to be used in agents, tooling layers, and coding assistants rather than only direct API prompting.

Pricing & Access

Published Mistral API pricing is:

  • Input: $0.40 per 1M tokens
  • Output: $2.00 per 1M tokens

Access paths include Mistral’s API and open-weight deployment routes. For teams running private infrastructure or regulated internal tooling, that open deployment path is a meaningful part of the model’s value.

Best Use Cases

Use Devstral 2 for coding agents, repo-aware CLI workflows, codebase modernization, and developer assistants where open weights and long context are both important. It is especially attractive when you want a Western open model that is explicitly optimized for software tasks instead of broad general chat.

It is less compelling as a single-model answer for everything. Teams should still pair it with stronger review discipline or a separate high-judgment reviewer model for risky migrations and security-sensitive work.

Comparisons

  • Mistral Small 4 (Mistral AI): Small 4 is the broader efficient open model; Devstral 2 is the coding-specialized route.
  • Qwen3.5 (Alibaba): Qwen3.5 is broader multilingual infrastructure; Devstral 2 is more intentionally positioned around developer workflows.
  • GPT-5.4 (OpenAI): GPT-5.4 is stronger as a proprietary frontier generalist; Devstral 2 is more interesting when openness, deployment control, and cost discipline matter.