Skip to main content
Browse Models

Mistral AI

Devstral Small 1.0

Released

2025-05-21

Family

Mistral

Type

Fine-Tuned Model

Huggingface PageOfficial Post

Downloads

External Download

You are about to open a link to an external source. Verify the URL before continuing.

Download

4-bit GGUF (Q4_K_M)

GGUF · mistralai_Devstral-Small-2505-Q4_K_M.gguf

5-bit GGUF (Q5_K_M)

GGUF · mistralai_Devstral-Small-2505-Q5_K_M.gguf

6-bit GGUF (Q6_K)

GGUF · mistralai_Devstral-Small-2505-Q6_K.gguf

8-bit GGUF (Q8_0)

GGUF · mistralai_Devstral-Small-2505-Q8_0.gguf

Model Report

Overview

Devstral is a specialized large language model (LLM) collaboratively developed by Mistral AI and All Hands AI, purpose-built for software engineering tasks. Finetuned from Mistral-Small-3.1, Devstral is engineered to act as an agentic coding assistant, with capabilities for codebase exploration, multi-file editing, and integration within software engineering agents. Released under the Apache 2.0 License on May 21, 2025, Devstral is available for both commercial and non-commercial uses and is positioned as a research preview to foster feedback and further development.

Agentic Performance / SWE-Bench Verified chart highlighting Devstral's performance-to-size ratio

Figure 1. Agentic Performance: Chart comparing Devstral with peer models on SWE-Bench Verified, showing its competitive performance relative to model size.

Full sizeSource

Technical Specifications and Architecture

Devstral is finetuned from Mistral-Small-3.1. It has a parameter count of 23.6 billion. The model features a context window of up to 128,000 tokens, allowing for extended reasoning and manipulation within large codebases. Devstral employs the Tekken tokenizer with a vocabulary size of 131,000 tokens, optimizing input and output handling for coding-oriented tasks. To focus exclusively on textual inputs relevant to software engineering, the vision encoder from its base model has been removed, making Devstral a text-only LLM.

The model’s architecture is tailored for agentic workflows. It enables contextual analysis across multiple files, identifies software component relationships, and assists in diagnosing code issues, which are functions for software engineering automation.

Training Methodology and Datasets

Devstral’s training regimen emphasizes real-world applicability to modern software engineering challenges. The model is trained to address GitHub issues using open-source agent scaffolds such as OpenHands and SWE-Agent. These scaffolds define interaction protocols between the model, codebases, and automated test cases, providing a framework for the model to learn how to effect multi-step code changes and verify correctness. Through this approach, Devstral is equipped to reason about, edit, and validate software repositories.

Benchmark Performance

Devstral demonstrates results in software engineering benchmarks, particularly on the SWE-Bench Verified evaluation. Utilizing the OpenHands scaffold for testing, Devstral achieves a score of 46.8%, outperforming previous open-source models by over 6 percentage points.

SWE-Bench benchmark scatter plot comparing model size and verified performance

Figure 2. SWE-Bench Benchmark Results: Devstral's performance compared to other models by size and verified solution rate.

Full sizeSource

Compared with similarly or larger-scaled models tested on the same scaffold, Devstral exceeds the performance of Claude 3.5 Haiku (40.6%), SWE-smith-LM 32B (40.2%), GPT-4.1-mini (23.6%), Deepseek-V3-0324 (38.8%, 671B), and Qwen3 232B-A22B (34.4%, 232B) models, as documented in benchmark summaries.

Usage Scenarios and Applications

Devstral is built for deployment within agentic software engineering systems. Its compact size facilitates local or on-device operation, which can be relevant for privacy-sensitive environments and integration into continuous integration/deployment workflows. Its primary applications include automated and interactive codebase analysis, bug detection, multi-file code editing, and test suite augmentation.

A common use case is repository test coverage analysis. Devstral can assess, aggregate, and visualize code test metrics, enabling identification of poorly covered modules and supporting targeted improvement.

Pie chart showing test coverage distribution output by Devstral-powered agent

Figure 3. Test Coverage Distribution: Devstral-generated pie chart of test coverage across files, produced in response to an analysis prompt.

Full sizeSource
Horizontal bar chart showing module-level test coverage produced by Devstral

Figure 4. Module-Level Detail: Horizontal bar chart of test coverage by module, produced by Devstral when prompted for detailed repository analysis.

Full sizeSource

In addition, Devstral can be embedded within agentic coding platforms and IDE plugins, functioning as a backend for interactive code completion, autonomous pull request generation, and resolution of multi-step tasks.

Minimalist To-Do List app UI generated by Devstral via OpenHands tutorial

Figure 5. To-Do List App Output: Minimalist application interface generated by Devstral as part of the OpenHands agent workflow tutorial.

Full sizeSource

Model Family and Future Development

Devstral is the first release in a planned family of agentic coding models. Finetuned from Mistral-Small-3.1, it is designed for compatibility with frameworks that define agent actions in software repositories, such as OpenHands. Mistral AI has indicated plans for future commercial and larger-scale agentic coding models featuring greater context lengths and domain adaptation, as outlined in the official announcement.

Limitations and Research Status

Devstral is released as a research preview, and continued improvement is anticipated through community feedback and operational experience. As with any LLM, its outputs are subject to the limitations of its training data and agent scaffold integration, and users are encouraged to validate and review changes proposed by the model in automated coding workflows.

Licensing

Devstral is distributed under the Apache 2.0 License, permitting use, modification, and distribution for both private and commercial applications.

Helpful Links

About Mistral: The Mistral family of AI models, developed by Paris-based Mistral AI, includes the original 2023 Mistral 7B release, as well as the more recent Mistral Small, Nemo, and Large weights.

More in the Mistral Family

Mistral AI /

Mistral Large 2

123 billion parameter model from Paris-based Mistral AI, significantly more capable than its predecessor in code generation, mathematics, reasoning, multilingual support, and function calling.
TheDrummer /

Behemoth 123B v1.2

A 123-billion parameter language model optimized for conversational AI, creative prose generation, and role-playing applications with enhanced narrative consistency.
Mistral AI /

Mistral Small (2409)

A 22B parameter enterprise-grade small model, a convenient mid-point between Mistral NeMo 12B and Mistral Large 2. This version delivers significant improvements in human alignment, reasoning capabilities, and code over the previous version.
Mistral AI /

Mistral Small 3.2 (2506)

A 24-billion parameter multimodal model featuring improved instruction following, function calling, and reduced repetition over its predecessor.
Mistral AI /

Mistral Small 3.1 (2503)

A 24-billion parameter multimodal transformer supporting text and vision tasks with 128K token context length under Apache 2.0 license.
LatitudeGames /

Harbinger 24B

A 24-billion parameter language model fine-tuned on Mistral Small 3.1 Instruct, specialized for interactive storytelling and text-based adventures.
Mistral AI /

Mistral Small 3 (2501)

A 24-billion parameter instruction-tuned language model with multilingual capabilities, 32K context window, and optimized low-latency inference performance.
TheDrummer /

Cydonia 24B v2

A fine-tuned 23.6 billion parameter Mistral-based model designed for long-context conversations and maintaining narrative coherence across extended dialogues.
Cognitive Computations /

Dolphin 3.0 Mistral 24B

A 24-billion parameter instruction-tuned model built on Mistral architecture with deliberately removed content filters to maximize user control over outputs.
Mistral AI /

Mistral NeMo 12B

A 12B parameter multi-lingual model that supports function calling built in collaboration with NVIDIA and trained using the new Tekken tokenizer. By some metrics, it is state-of-the-art in its size category. NeMo was trained with quantisation awareness, enabling FP8 inference without any performance loss.
TheDrummer /

Rocinante 12B v1.1

A 12.2 billion parameter text generation model optimized for creative storytelling, role-playing scenarios, and adventure-based interactive fiction applications.

More from Mistral AI

Mistral AI /

Mistral 7B

Base 7.3-billion parameter model from Mistral AI, with superior performance to Llama 2 7b (and 13b, impressively) on nearly all benchmarks.
Mistral AI /

Codestral 22B v0.1

A 22.2 billion parameter causal language model trained on over 80 programming languages with support for code generation and fill-in-the-middle completion.
Mistral AI /

Mixtral 8x7B

Spare sparse mixture of experts (MoE) model consisting of 8 "experts" with 7b parameters each, outperforming all Llama 2 models (including 70b) on most benchmarks.
Mistral AI /

Mixtral 8x22B

A sparse mixture-of-experts architecture with 141 billion total parameters, activating 39 billion per token, supporting multilingual tasks and 64,000-token contexts.

Compatible Apps

Open WebUI
open-webui /

Open WebUI

A polished, self-hosted chat interface for LLMs with Ollama integration, multimodal prompts, and extensive workspace customization.

Web UI

Chat UIs · Beginner Friendly

L
ggml-org /

llama.cpp

GGUF model inference with a polished web UI and OpenAI-format API. CPU-only build — high hardware compatibility, works on any machine without a GPU.

Web UI · API · CLI

LLM Inference · Chat UIs

L
ggml-org /

llama.cpp (CUDA)

GGUF model inference with a polished web UI and OpenAI-format API. CUDA build — GPU-accelerated for NVIDIA GPUs.

Web UI · API · CLI

LLM Inference · Chat UIs

Text Generation Web UI
oobabooga /

Text Generation Web UI

A feature-rich interface for running and experimenting with open-weight LLMs, including multiple inference backends, plugins, and tuning controls.

Web UI · API

Chat UIs · LLM Inference