Qwen 2.5-Max models by Alibaba are better than DeepSeek, or so it claimed

The US market is shaking up as Alibaba, an Amazon rival from China, is entering the AI boom with a model that it claims to be more powerful than the now-surging DeepSeek.

The tech giant has now released its Qwen 2.5-Max model, claiming it outperforms DeepSeek-V3, GPT-4o, and Llama-3.1-405B across multiple benchmarks. The release, timed on the first day of the Lunar New Year, comes amid growing pressure from the rapid rise of DeepSeek.

Qwen 2.5-Max, available in 3B, 7B, and 72B parameter sizes, is also open-source, with versions released on Hugging Face and ModelScope. It’s also available on Alibaba Cloud and the Qwen Chat interface and offers an OpenAI-compatible API.

The model’s training included 18-20 trillion tokens and used dynamic resolution/frame-rate techniques, which improves its efficiency through multi-stage RLHF and SFT with over 1 million samples.

Benchmark tests reveal that Qwen 2.5-Max not only excels in AI tasks but also dominates in coding and reasoning with big reductions in resource usage. It surpasses DeepSeek V3 in benchmarks like Arena-Hard, LiveBench, and LiveCodeBench, while also being competitive in others such as MMLU-Pro.

DeepSeek’s arrival in the AI scene caused a massive hysteria in the US market, with at least over $1 trillion loss in the US stock market. Both Nvidia and Microsoft are among American tech firms that were affected as competitors from China are looking strong.

But the challenge doesn’t only come from outside. In China, companies are also forced to lower their AI prices, especially after DeepSeek claimed that it’s built its ChatGPT rival with just $6 million and older Nvidia H100 GPUs.