Atlas

Models

← All models

Devstral Small 2

Mistral AI · 2025-12-09 · 24B parameters

Devstral Small 2 is Mistral AI's December 2025, 24B-class open-weight multimodal agentic-coding model, released alongside the 123B Devstral 2. It accepts text and image inputs and is designed for tool-driven codebase exploration, multi-file editing, and software-engineering agents. The checkpoint supports a 262,144-token context, is published in FP8 form, and carries the Apache-2.0 weight license.

Benchmark scores

BenchmarkScore
AA-LCR24.0
AA-Omniscience Index-56.7
AIME 202534.3
Artificial Analysis Agentic Index10.5
Artificial Analysis Coding Index29.3
Artificial Analysis Intelligence Index17.4
Artificial Analysis Omniscience Accuracy15.3
Artificial Analysis Omniscience Hallucination Rate85.1
Artificial Analysis Openness Index27.8
Capability128.8
CritPt0.0
GDPval-AA v2734
GDPval-AA v2 (normalized)11.7
Global MMLU Lite69.2
GPQA Diamond53.2
Humanity's Last Exam (Text-Only)3.4
IFBench (Artificial Analysis)31.2
IOI2.5
IOI - IOI 20243.7
IOI - IOI 20251.3
LiveCodeBench (Artificial Analysis)34.8
LiveCodeBench (Vals AI public dataset)43.2
LiveCodeBench (Vals Public) - Easy85.7
LiveCodeBench (Vals Public) - Hard8.6
LiveCodeBench (Vals Public) - Medium35.2
MMLU Pro67.8
MMMU Pro44.6
SciCode28.8
SciCode (Artificial Analysis)28.8
Terminal-Bench 2.129.6
Loading Atlas data…