Atlas

Models

← All models

Devstral 2

Mistral AI · 2025-12-09 · 125B parameters

Devstral 2 is Mistral AI's dense, text-only, 123B-class open-weight agentic coding model, released as the larger member of the Devstral 2 family alongside Devstral Small 2. Its instruction-tuned FP8 checkpoint supports 262,144 tokens through YaRN extrapolation from an original 4,096-position setting and is designed for tool-driven codebase exploration, multi-file editing, and software-engineering agents.

Benchmark scores

BenchmarkScore
AA-LCR30.0
AA-Omniscience Index-46.2
AIME 2025 (Artificial Analysis avg@10)36.7
Artificial Analysis Agentic Index10.6
Artificial Analysis Coding Index31.3
Artificial Analysis Intelligence Index19.2
Artificial Analysis Omniscience Accuracy21.0
Artificial Analysis Omniscience Hallucination Rate85.1
Artificial Analysis Openness Index27.8
BuseyBench SVG2.4
Capability136.0
CritPt (avg@5)0.0
GBENCH One-shot Coding GScore23.5
GDPval-AA v2745
GDPval-AA v2 (normalized)12.3
Global MMLU Lite77.5
GPQA Diamond (Artificial Analysis avg@5)59.4
Humanity's Last Exam (Text-Only)3.6
IFBench (Artificial Analysis)38.1
IOI0.7
IOI - IOI 20240.0
IOI - IOI 20251.3
LiveCodeBench (Artificial Analysis)44.8
LiveCodeBench (Vals AI public dataset)51.8
LiveCodeBench (Vals Public) - Easy90.7
LiveCodeBench (Vals Public) - Hard13.1
LiveCodeBench (Vals Public) - Medium51.7
MMLU Pro76.2
ObviousBench65.3
Pencil Puzzle Bench - Agentic0.0
Loading Atlas data…