Overview
Devstral Small, released May 21, 2025, was the first model in Mistral AI's Devstral line — an agentic large language model for software engineering, built in collaboration with All Hands AI. It is a 24-billion-parameter open model under the Apache 2.0 license, designed to solve real GitHub issues rather than answer isolated coding prompts.
On SWE-Bench Verified, Devstral Small scores 46.8%, which Mistral reported as outperforming prior open-source state-of-the-art by about 6 points and, under the same OpenHands scaffold, exceeding far larger models such as DeepSeek-V3-0324 (671B) and Qwen3 235B-A22B. It uses a 128K-token context window and the Tekken tokenizer with a 131k-token vocabulary.
Devstral Small is fine-tuned from Mistral-Small-3.1-24B-Base-2503 and is text-only — the vision encoder was removed for coding-agent use. It is light enough to run on a single RTX 4090 or a 32GB Mac, and was distributed through the Mistral API (model ID devstral-small-2505), Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio. It has since been superseded by the Devstral 1.1 and Devstral 2 generations.
| Released | 2025-05-21 |
|---|---|
| License | Apache 2.0 |
| Weights | Open weights |
| Parameters | 24B |
| Context | 128K |
| Architecture | Dense decoder-only Transformer; fine-tuned from Mistral-Small-3.1-24B-Base-2503 (text-only, vision encoder removed); Tekken tokenizer (131k vocab) |
| Knowledge cutoff | 2025 |
| Modalities | Text |
| Status | Deprecated |
Benchmarks
- SWE-Bench Verified46.8%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.10 / 1M tokens per 1M tokens |
|---|---|
| Output | $0.30 / 1M tokens per 1M tokens |
Mistral API pricing as published in the May 2025 announcement; weights are also free under Apache 2.0.
Strengths
- First open Apache 2.0 Devstral coding agent — fully self-hostable
- Runs locally on a single RTX 4090 or a 32GB Mac
- Outperformed much larger open models on SWE-Bench Verified under the same OpenHands scaffold
- 128K-token context for repository-scale agentic tasks
- Widely distributed: Hugging Face, Ollama, Kaggle, Unsloth, LM Studio, and the Mistral API
Best for
- Open, self-hosted coding agents for resolving GitHub issues
- Local, privacy-preserving software-engineering assistants on a single GPU
- Agentic code workflows inside OpenHands-style scaffolds
- Cost-efficient code generation through the Mistral API
How to access
| Provider | Model ID |
|---|---|
| Mistral AI ↗ | devstral-small-2505 |
Devstral — every version
The full lineage of the Devstral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Devstral 2 / Devstral Small 2current | 2025-12-09 | 256K | Devstral 2: Modified MIT; Devstral Small 2: Apache 2.0 |
| Devstral Medium & Small 1.1 (25.07) | 2025-07-10 | 128K | Devstral Small 1.1: Apache 2.0; Devstral Medium: proprietary (API only) |
| Devstral Small (25.05) | 2025-05-21 | 128K | Apache-2.0 |
FAQ
What is Devstral Small (25.05)?
Devstral Small is Mistral AI's first Devstral model, released May 21, 2025 with All Hands AI. It is a 24B open Apache 2.0 agentic coding model designed to solve real GitHub issues, scoring 46.8% on SWE-Bench Verified with a 128K context window.
Is Devstral Small open source and can it run locally?
Yes. It is released under Apache 2.0 with open weights on Hugging Face (model ID devstral-small-2505), fine-tuned from Mistral-Small-3.1-24B-Base-2503. It is light enough to run on a single RTX 4090 or a Mac with 32GB RAM and is also available via Ollama, Kaggle, Unsloth, and LM Studio.
How did Devstral Small compare to larger models?
Mistral reported that, under the same OpenHands scaffold, Devstral Small's 46.8% SWE-Bench Verified score exceeded much larger open models such as DeepSeek-V3-0324 (671B) and Qwen3 235B-A22B, while outperforming the prior open-source state-of-the-art by about 6 points.
Is Devstral Small still the current version?
No. Devstral Small (25.05) has been superseded by Devstral Small 1.1 (July 2025) and the December 2025 Devstral 2 family. Its Apache 2.0 weights remain available on Hugging Face for download and self-hosting.