Qwen2.5-Max challenges DeepSeek 🧠China’s ‘artificial sun’ ☀️ Solar farm in space 🚀

2025-02-03 | subscribe | homepage

Welcome back to this week’s edition of the China Insights Weekly Newsletter!

Key highlights shaping the week:

• Alibaba’s Qwen2.5-Max leads AI benchmarks with 20T tokens trained

• DeepSeek app tops iPhone downloads across US and Europe

• Artificial sun sets fusion record with 1,066 seconds of plasma

• China renewables hit 56% capacity, outpacing residential demand

Dive deeper into these stories and more by clicking the headlines below. We value your feedback. Let us know your thoughts or suggestions on LinkedIn, X or Facebook.

🚀 Headlines

Alibaba’s Qwen2.5-Max outperforms DeepSeek and other major AI models on major benchmarks (link)

Qwen2.5-Max, a large-scale Mixture-of-Expert (MoE) model developed by Alibaba’s Qwen team, has been trained on over 20 trillion tokens and further refined with Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). The model demonstrates superior performance in various benchmarks, outperforming DeepSeek V3 in Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, and showing competitive results in MMLU-Pro. Qwen2.5-Max is now available via Qwen Chat and its API, which is accessible through Alibaba Cloud. Following DeepSeek and the new Qwen models, the competition out of China for open-source AI models continues to increase.

DeepSeek AI assistant app tops iPhone downloads for a full week in the US and major Western markets (link)

DeepSeek, a Chinese AI assistant app, surged to the top of Apple's App Store, surpassing OpenAI's ChatGPT. It has been the most downloaded free app in the US and major Western markets such as the UK, Germany, and Spain for over the past week. The app is known for its fast and logical responses, with some users claiming its natural language processing and reasoning capabilities outperform those of its US counterparts. DeepSeek was founded by Chinese hedge fund manager Liang Wenfeng and launched its AI model DeepSeek R1 at the World Economic Forum in Davos. The model was developed in a few months for about USD 6 million using 2,000 H800 Nvidia chips, which have lower data transfer rates compared to advanced US chips.