Atlas

Models

← All models

Qwen3 30B A3B Instruct 2507

Alibaba · 2025-07-30 · 30.5B parameters

Qwen3-30B-A3B-Instruct-2507 is Alibaba's non-thinking July 2025 instruction update to Qwen3-30B-A3B. Its checkpoint contains exactly 30,532,122,624 parameters, while Qwen reports 3.3B activated only at rounded precision. The model has a native 262,144-token context window and the published DCA/MInference configuration supports an exact 1,010,000-token serving length; it is distinct from the Thinking checkpoint.

Benchmark scores

BenchmarkScore
AA-LCR22.7
AA-LCR v1.126.3
AA-Omniscience Index-66.2
AIME 2025 (Artificial Analysis avg@10)66.3
Artificial Analysis Omniscience Accuracy15.0
Artificial Analysis Omniscience Hallucination Rate95.4
Artificial Analysis Openness Index44.4
Artificial Analysis Progress Index v1199
BFCL v4 Overall Accuracy41.4
BuseyBench SVG2.6
Capability129.4
Chess Puzzles2.0
CritPt (avg@5)0.0
DTBench67.2
Epoch Capabilities Index (ECI)137
Global MMLU Lite72.0
GPQA Diamond55.6
GPQA Diamond (Artificial Analysis avg@5)65.9
Humanity's Last Exam (Text-Only)6.9
IFBench (Artificial Analysis avg@5)33.1
Judgemark v2.134.6
LisanBench Difficulty-Weighted19.0
LisanBench Path Length74.7
LiveCodeBench (Artificial Analysis)51.5
LMCA22.4
MATH-500 (Artificial Analysis source)97.5
MMLU Pro77.7
ObviousBench50.0
OTIS Mock AIME 2024-202562.2
SciCode (Artificial Analysis avg@3)30.4
Loading Atlas data…