Qwen2.5-Max
Alibaba's large MoE flagship, pretrained on over 20T tokens, served by API and Qwen Chat; claims wins over DeepSeek V3 on several benchmarks.
Large-scale MoE with SFT and RLHF, announced eight days after R1. Alibaba says it beats DeepSeek V3 on Arena-Hard, LiveBench, LiveCodeBench and GPQA-Diamond. The blog states many scaling details were only disclosed with DeepSeek V3's release.
- Date
- Tuesday, 28 January 2025
- Lab
- Alibaba (Qwen)
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Pretraining tokens | over 20T Large MoE; API through Alibaba Cloud and Qwen Chat; weights closed | company |
Blog dated 2025-01-28. Parameter count and pricing not disclosed in the post. Benchmark comparisons are Alibaba's own and sometimes against base models; the nod to DeepSeek V3's disclosures is a direct example of cross-lab diffusion.
Sources
This record was checked against its sources on 6 October 2026. How we check