Atlas

Models

← All models

Qwen3-Max-Preview (Thinking Mode)

Alibaba · 2025-11-03

Alibaba made thinking mode available for the hosted qwen3-max-preview model on November 3, 2025, improving its reasoning and agentic-programming capabilities. The model used the shared qwen3-max-preview API ID with hybrid thinking controlled by enable_thinking; Alibaba later released the separate qwen3-max-2026-01-23 snapshot.

Benchmark scores

BenchmarkScore
AA-LCR57.7
AA-Omniscience Index-37.8
AIME 2025 (Artificial Analysis avg@10)82.3
Artificial Analysis Intelligence Index25.0
Artificial Analysis Omniscience Accuracy27.9
Artificial Analysis Omniscience Hallucination Rate91.3
Artificial Analysis Openness Index16.7
Capability141.7
CritPt (avg@5)0.0
Elimination Game TrueSkill Mu4.5
Global MMLU Lite85.5
GPQA Diamond (Artificial Analysis avg@5)77.6
Humanity's Last Exam (Text-Only)12.0
IFBench (Artificial Analysis)53.8
LiveCodeBench (Artificial Analysis)53.5
MineBench Conservative Rating854
MMLU Pro82.4
SciCode (Artificial Analysis avg@3)38.7
Step Game TrueSkill Mu1.5
Terminal-Bench Hard17.4
WordleBench86.0
WordleBench — Average Guesses on Successful Games3.7
τ²-bench (Telecom)83.6
Loading Atlas data…