## 🔍 Key Insights Global weekly token usage for large models has reached **211 trillion**, with domestic models leading the US for 29 consecutive weeks at **119.1 trillion** [7]. Meanwhile, **Qwen4-Max** has set a new flagship specification with a **2.4-trillion-parameter MoE** architecture [12], and **Perplexity** is investing at least **10,000 GPUs** to enter the automated researcher track [3]. The arms race in model capabilities and computing power is heating up in tandem. ## 🚀 Top Developments - **Qwen releases Qwen4-Max** [12]: A 2.4-trillion-parameter MoE flagship with native visual understanding and a context window of 1 million tokens. - **Perplexity launches Frontier Research** [3]: Investing at least 10,000 GPUs to research multi-agent orchestration and continual learning. - **OpenRouter: Domestic models lead for 29 consecutive weeks** [7]: Global token usage last week reached 211 trillion, with Chinese models at 119.1 trillion. - **US tightens AI chip export controls** [5]: AI chips and servers exported to global data centers now require licenses, closing off remote computing power access. - **Liquid AI releases d1.2 decision models** [4]: d1.2-ultra tops the AA-Paperclip benchmark with 76 points, leading the second place by 3 points. - **Tencent Hy4 flash hands-on test** [1]: For just 0.2 yuan, AI can build a 3D interactive webpage from scratch, showing outstanding cost-effectiveness for high-frequency productivity scenarios. - **Google Maps launches Ask Maps trip planning** [17]: Powered by Gemini, it generates day-by-day itineraries and supports export to Google Docs, expanding to 30 European countries [16]. - **Apple smart glasses N50 details revealed** [2]: Screenless, voice-interaction-first, planned for WWDC 2027 launch, relying on Siri and the ecosystem to fill the gap. ## 🔗 Sources [1] Hands-on test of Tencent Hy4 flash: For just 0.2 yuan, AI can build a 3D interactive webpage from scratch — https://www.bestblogs.dev/article/4e2dd28fd9?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [2] Apple smart glasses details revealed — they're basically AirPods worn on your face — https://www.bestblogs.dev/article/2d31435f58?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [3] Perplexity launches Frontier Research, investing at least 10,000 GPUs to build an automated researcher — https://aihot.news/items/qx8tl0o0zfzb169u5frk96o0e [4] Liquid AI releases d1.2-ultra and d1.2 decision models — https://aihot.news/items/c7vgnvk5g20kcz9cu90jxp4gq [5] US tightens AI chip export controls, restricting Chinese companies' remote access to computing power — https://aihot.news/items/x6msqkehvwzry0904un629cqg [7] OpenRouter data: Global large model token usage reached 211 trillion last week, Chinese models lead US for 29 consecutive weeks — https://aihot.news/items/zgnjy5wmajgdqzpt0ds98o2qx [12] Qwen releases Qwen4-Max, a 2.4-trillion-parameter MoE flagship model — https://aihot.news/items/dnudjlqdqlc6c9h3