Atlas

Models

← All models

Qwen2.5-Max

Alibaba · 2025-01-27

Qwen2.5-Max is Alibaba's large-scale closed-weight MoE flagship in the Qwen2.5 family, released in Alibaba Cloud Model Studio on January 27, 2025 and announced by Qwen on January 28. It was pretrained on over 20 trillion tokens, then post-trained with curated SFT and RLHF, and was offered through the Model Studio API and Qwen Chat. Alibaba retired the `qwen-max-2025-01-25` snapshot from Model Studio on May 13, 2026.

Benchmark scores

BenchmarkScore
Aider Polyglot21.8
Artificial Analysis Intelligence Index10.2
BigCodeBench-Hard Complete35.8
BigCodeBench-Hard Instruct29.1
Capability129.9
Confabulations Leaderboard Confabulation Rate31.2
Confabulations Leaderboard Non-Response Rate12.4
Confabulations Leaderboard Weighted Score21.8
Elimination Game TrueSkill Mu2.5
Epoch Capabilities Index (ECI)133
Fiction.liveBench66.7
Humanity's Last Exam (Text-Only)4.5
LiveBench62.3
LiveCodeBench (Artificial Analysis)35.9
LLM Creative Story-Writing Benchmark7.3
MATH Level 567.2
MATH-500 (Artificial Analysis source)83.5
MMLU Pro76.2
OTIS Mock AIME 2024-202516.1
PGG-Bench TrueSkill Mu9.5
Phare Average Safety62.2
Phare Bias Resistance43.0
Phare Debunking85.4
Phare Debunking (English)87.3
Phare Debunking (French)85.6
Phare Debunking (Spanish)83.3
Phare Encoding Jailbreaks27.6
Phare Encoding Jailbreaks (English)27.6
Phare Encoding Jailbreaks (French)33.1
Phare Encoding Jailbreaks (Spanish)22.3
Loading Atlas data…