Devstral

What Devstral is

Devstral is a family of coding models from Mistral AI, released on 21 May 2025 as a research preview and built together with All Hands AI, the company behind the OpenHands coding agent^[1]. Its selling point is not autocomplete but "agentic coding": the model is meant to work inside a loop, looking at a repository, deciding which files to open, editing them and calling tools, while an external scaffold handles the mechanics and feeds results back.

The problem it targets

The benchmark Mistral used to place the model is SWE-Bench Verified, which gives a system real issues from real open-source repositories and checks whether the resulting code change resolves them. The first Devstral scored 46.8% on SWE-Bench Verified, which Mistral described as more than 6 points above the previous best open-source result^[1]. Mistral also reported that, running inside the OpenHands scaffold, it beat DeepSeek-V3-0324 (671B parameters) and Qwen3 232B-A22B^[1]. That comparison is the interesting part: both are models with hundreds of billions of parameters, while Devstral was built to fit on ordinary hardware.

Open weights and a small footprint

Devstral was released under the Apache 2.0 licence and can run on a single RTX 4090 or a 32GB Mac^[1]. Through Mistral's API it was priced at $0.10 per million input tokens and $0.30 per million output tokens under the model name devstral-small-2505^[1]. Mistral said the model runs over code agent scaffolds such as OpenHands or SWE-Agent^[1], a reminder that a coding model is only one component of such a system: how the scaffold assembles context, chooses files and invokes tools matters as much as the model itself.

The July 2025 second generation

On 10 July 2025 Mistral released Devstral Small 1.1 and Devstral Medium^[2]. Small 1.1 has 24B parameters, stays under Apache 2.0, and scored 53.6% on SWE-Bench Verified; Medium reached 61.6% and was offered only through Mistral's public API, at $0.40 per million input tokens and $2 per million output tokens^[2]. Mistral said the Devstral 2507 models support both Mistral's function calling format and XML formats^[2]. In practice that means the model can express a tool call either as a structured function-call object or as XML markup, so it can be dropped into scaffolds written for either convention without rewriting the interface.

Retirement and successor

Devstral Small 1.0 was retired on 30 November 2025, and Devstral Small 1.1 and Devstral Medium on 31 May 2026^[3]. The line was succeeded by Devstral 2 in December 2025^[3].

Where it sits

Devstral matters for two reasons. First, it was an explicit test of whether a model small enough to run on a single consumer GPU could be competitive at agentic coding: the reported SWE-Bench Verified score and the head-to-head comparison with far larger models were the evidence Mistral offered^[1], and the later Small 1.1 and Medium releases pushed the same benchmark higher while keeping one version under an open licence^[2]. Second, the release pattern is informative. Mistral shipped an open-weight model and an API model side by side, priced both per token, and retired the first version within months of launch^[1][3] — a pace that says as much about how quickly agentic coding models were being revised as about the model itself.