Atlas

Models

← All models

Devstral Small 1.1

Mistral AI · 2025-07-10 · 23.6B parameters

Devstral Small 1.1 is an open-weight 24B agentic coding model co-developed by Mistral AI and All Hands AI, fine-tuned from Mistral Small 3.1 after its vision encoder was removed. It is text-only, supports a 131,072-token context, and is designed for codebase exploration, multi-file editing, tool use, function calling, and varied coding-agent scaffolds. Mistral released the weights under Apache-2.0 and lists the hosted API snapshot as deprecated on February 27, 2026, with Mistral Small 4 as its replacement.

Benchmark scores

BenchmarkScore
AA-LCR17.0
AA-Omniscience Index-51.1
AIME 202529.3
Artificial Analysis Intelligence Index9.3
Artificial Analysis Omniscience Accuracy14.7
Artificial Analysis Omniscience Hallucination Rate77.1
Artificial Analysis Openness Index44.4
Capability120.2
CritPt0.0
Design Arena Overall Code Categories837
EDIT-Bench Complete36.7
EDIT-Bench Complete Easy60.3
EDIT-Bench Complete Hard8.2
EDIT-Bench Core48.1
EDIT-Bench Core Easy72.6
EDIT-Bench Core Hard15.2
GPQA Diamond41.4
Humanity's Last Exam (Text-Only)3.7
IFBench (Artificial Analysis)34.6
LiveCodeBench (Artificial Analysis)25.4
MATH-500 (Artificial Analysis source)63.5
SciCode (Artificial Analysis)24.3
SnakeBench Average Apples2.5
SnakeBench Best Apples9.0
SnakeBench Rating25.3
SnakeBench Total Apples67.0
SnakeBench Win Rate43.5
Terminal-Bench Hard6.1
τ²-bench (Telecom)28.4
Loading Atlas data…