Devstral Small

基於 Mistral-Small-3.1 微調的 24B 參數開源程式碼智能體模型，專為軟體工程任務設計。在 SWE-Bench Verified 上達到 53.6%（v1.1 版本），支援 128K 上下文視窗。可在單個 RTX 4090 或 32GB 記憶體 Mac 上運行，採用 Apache 2.0 許可證，專為 OpenHands 和 SWE-Agent 等程式碼智能體框架優化。

Intelligence(Relatively Weak)

Speed(Medium)

Input Supported Modalities

Is Reasoning Model

256,000

Context Window

128,000

Maximum Output Tokens

Knowledge Cutoff

Go Compare

Pricing

￥0.72 /M tokens

Input

￥2.16 /M tokens

Output

￥1.08 /M tokens

Blended Price

Quick Simple Comparison

Input

Output

Devstral Small (May '25)

￥0.1

Codestral (Jan '25)

￥0.2

Devstral Small

￥0.1

Basic Parameters

Devstral SmallTechnical Parameters

Parameter Count

24,000.0M

Context Length

256.00k tokens

Training Data Cutoff

Open Source Category

Open Weights (Permissive License)

Multimodal Support

Text Only

Throughput

137

Release Date

2025-07-10

Response Speed

141.78,554 tokens/s

Benchmark Scores

Below is the performance of Devstral Small in various standard benchmark tests. These tests evaluate the model's capabilities in different tasks and domains.

Intelligence Index

Large Language Model Intelligence Level

32.07

Large Language Model Intelligence Level

Coding Index

Indicator of AI model performance on coding tasks

24.85

Indicator of AI model performance on coding tasks

Math Index

Capability indicator in solving mathematical problems, mathematical reasoning, or performing math-related tasks

31.9

Capability indicator in solving mathematical problems, mathematical reasoning, or performing math-related tasks

MMLU Pro

Massive Multitask Multimodal Understanding - Testing understanding of text, images, audio, and video

62.2

Massive Multitask Multimodal Understanding - Testing understanding of text, images, audio, and video

GPQA

Graduate Physics Questions Assessment - Testing advanced physics knowledge with diamond science-level questions

41.4

Graduate Physics Questions Assessment - Testing advanced physics knowledge with diamond science-level questions

HLE

The model's comprehensive average score on the Hugging Face Open LLM Leaderboard

3.7

The model's comprehensive average score on the Hugging Face Open LLM Leaderboard

LiveCodeBench

Specific evaluation focused on assessing large language models' ability in real-world code writing and solving programming competition problems

25.4

Specific evaluation focused on assessing large language models' ability in real-world code writing and solving programming competition problems

SciCode

The model's capability in code generation for scientific computing or specific scientific domains

24.3

The model's capability in code generation for scientific computing or specific scientific domains

HumanEval

Score achieved by the AI model on the specific HumanEval benchmark test set

Math 500 Score

Score on the first 500 larger, more well-known mathematical benchmark tests

63.5

Score on the first 500 larger, more well-known mathematical benchmark tests

AIME Score

An indicator measuring an AI model's ability to solve high-difficulty mathematical competition problems (specifically AIME level)

0.3

An indicator measuring an AI model's ability to solve high-difficulty mathematical competition problems (specifically AIME level)

Output tokens/million

Output tokens/million

Output tokens/million

Output tokens/million

Output tokens/million

131k

Context Length

Gemini 2.0 Flash Lite (Preview)

Output tokens/million

Output tokens/million

Output tokens/million

8192

Context Length

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご