Alibaba Announces "Qwen2.5-Max" — an MoE Model It Claims Beats DeepSeek-V3

On January 28, 2025, Alibaba announced its large-scale Mixture-of-Experts model, “Qwen2.5-Max.” Launched right after DeepSeek shook the world with V3 and R1, Alibaba says the model outperforms DeepSeek-V3 on benchmarks including Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, with performance on some metrics matching or exceeding GPT-4o.

Details

• Architecture: A large-scale Mixture-of-Experts model pretrained on more than 20 trillion tokens, then refined with supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF)
• Benchmarks: Alibaba claims the model surpasses DeepSeek-V3 on Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, and achieves competitive results on MMLU-Pro and other benchmarks
• Context: The announcement came right after DeepSeek’s V3 and R1 shook the industry, symbolizing the intensifying AI development race within China
• Availability: Released as an API on Alibaba Cloud, with API access also promoted through Qwen Chat

What happened next

Qwen2.5-Max was a move that demonstrated Alibaba’s ability to compete with the world’s top closed models even at the flagship level. This trajectory continued into “Qwen3” in April 2025, followed by the rapid succession of Qwen3.5, Qwen3.6, and Qwen3.7 starting in 2026.