Atlas

Models

← All models

Qwen3 Max Thinking (2026-01-23)

Alibaba · 2026-01-27

Qwen3-Max-2026-01-23 is the hosted API snapshot of Alibaba's closed-weight Qwen3-Max-Thinking flagship reasoning model, publicly released on January 27, 2026. It supports hybrid thinking and non-thinking modes.

Benchmark scores

BenchmarkScore
AA-LCR66.0
AA-LCR v1.174.3
AA-Omniscience Index-36.6
Artificial Analysis Omniscience Accuracy29.8
Artificial Analysis Omniscience Hallucination Rate94.7
Artificial Analysis Progress Index v1265
BullshitBench v2 Leaderboard63.0
Capability145.8
CaseLaw55.0
CorpFin v268.0
CorpFin v2 - Exact Pages64.9
CorpFin v2 - Max Fitting Context70.6
CorpFin v2 - Shared Max Context68.5
CritPt (avg@5)1.7
Eval Connections95.0
Eval Connections — Guess Accuracy89.7
Eval Connections One-Shot Group Accuracy53.8
Eval Connections One-Shot Puzzle Solves50.0
Eval Connections One-Shot Trap Bonus5.0
Eval Connections One-Shot with Traps35.0
Extended NYT Connections 940 (Quadratic v1)30.1
Extended NYT Connections 940-Puzzle Set42.1
Extended NYT Connections Newest 10042.5
Extended NYT Connections Newest 100 (Quadratic v1)31.6
GPQA Diamond86.9
GPQA Diamond (Artificial Analysis avg@5)86.1
GPQA Diamond (Vals protocol average)84.8
Humanity's Last Exam (Text-Only)28.0
IFBench (Artificial Analysis avg@5)70.7
IOI13.8
Loading Atlas data…