Atlas

Models

← All models

Qwen3 235B A22B Thinking 2507

Alibaba · 2025-07-25 · 235.1B parameters

Qwen3-235B-A22B-Thinking-2507 is Alibaba's July 25, 2025 thinking-only update to Qwen3-235B-A22B. The released checkpoint contains 235,093,634,560 parameters; Qwen reports 22B activated parameters only at rounded precision. It supports 262,144 tokens natively, with an optional DCA/MInference configuration documented at a 1,010,000-token runtime limit.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo190
AA-Briefcase v1.1 Elo170
AA-LCR67.0
AA-LCR v1.172.0
AA-Omniscience Index-46.5
AIME 2025 (Artificial Analysis avg@10)91.0
Apex (MathArena)5.2
Artificial Analysis Agentic Index3.8
Artificial Analysis Coding Index22.1
Artificial Analysis Intelligence Index19.6
Artificial Analysis Intelligence Index v4.312.7
Artificial Analysis Intelligence Index v4.3.212.7
Artificial Analysis Omniscience Accuracy22.8
Artificial Analysis Omniscience Hallucination Rate89.8
Artificial Analysis Openness Index44.4
Artificial Analysis Progress Index v1237
AutomationBench-AA 1.0.62.2
BuseyBench SVG2.4
Capability141.5
Chess Puzzles12.0
Confabulations Leaderboard Confabulation Rate13.9
Confabulations Leaderboard Non-Response Rate19.7
Confabulations Leaderboard Weighted Score16.8
CritPt (avg@5)0.0
Design Arena Overall Code Categories1031
DTBench80.3
Epoch Capabilities Index (ECI)144
Fiction.LiveBench - 0 tokens (relevant passages only)100.0
Fiction.LiveBench - 120K tokens68.8
Fiction.LiveBench - 16K tokens75.0
Loading Atlas data…