Devstral 2

What it is

Devstral 2 is an agentic coding model released by Mistral AI on 9 December 2025, together with a command-line tool called Mistral Vibe CLI ^[1]. Agentic here means the model is built to run in a loop inside a tool rather than to answer a single question: it reads files, proposes edits, runs commands, looks at what came back and tries again. Working this way means holding a lot of material in view at once — source files, error output, earlier attempts — so the size of the context window matters as much as raw capability.

Devstral 2 is a dense Transformer with 123 billion parameters and a context window of 256K tokens ^[1]. Dense means every parameter is used for every token, in contrast with mixture-of-experts designs that route each token through only part of the network. A window of 256K tokens is on the order of several hundred pages of text, enough to hold the relevant files of a mid-sized project plus the running conversation.

Two sizes

The release came in two sizes. Devstral Small 2 has 24 billion parameters, the same 256K context window, a score of 68.0% on SWE-bench Verified, and is released under the Apache 2.0 licence ^[1]. The larger Devstral 2 scored 72.2% on SWE-bench Verified ^[1].

SWE-bench Verified is a benchmark assembled from real issue reports in open-source Python projects. A model is given a repository and an issue and must produce a patch that makes the project's tests pass. Grading is automatic, which is why the benchmark is widely used to compare coding models, and also why it captures only part of what programming involves.

Licence and access

Devstral 2 was released under a modified MIT licence ^[1]. MIT is one of the most permissive standard open-source licences; the modified qualifier means conditions differ from the standard text, so the terms have to be read in the licence itself.

On the API side the model is listed under the name devstral-2512 ^[3]. Access was free for a period, after which the published prices were $0.40 per million input tokens and $2.00 per million output tokens for Devstral 2, and $0.10 and $0.30 respectively for Devstral Small 2 ^[1].

Timeline

Mistral released Mistral Vibe 2.0 on 27 January 2026, and Devstral 2 moved to paid API access the same day ^[2][3]. The model was marked deprecated in Mistral's models overview on 22 May 2026 ^[3], and retired on 31 July 2026, when it was replaced by Mistral Medium 3.5 ^[3].

Where it sits

From release to retirement was under eight months. Devstral 2 arrived as a package — a model, a command-line tool to drive it, a cheaper sibling released at the same time, and published per-token prices — aimed at code work done by an agent rather than through a chat window. Its replacement carries a different name, Mistral Medium 3.5 ^[3], so the flagship of the Devstral line did not continue under the same label, and anyone calling devstral-2512 had to move to another model once the API entry was retired ^[3].