Hi~~ Good evening, everyone! Today is Sunday, October 4, 2026. I’m An’an, host of Goodnight Coffee, and I’m right on time to meet you again! Tonight, let’s talk about the freshest business and tech developments and see what big things are happening. So relax, bring a good mood, and let’s get started! @@@ Sina Tech reports that Huawei held a launch event on October 1 for the Mate 90 series and all-scenario new products, with Yu Chengdong, Huawei Executive Director and Chairman of the Terminal BG, delivering the keynote. This is the first time Huawei’s flagship phones have launched and gone on sale on October 1. The lineup includes the Mate 90, Pro, Pro Max, Pro Max Collector’s Edition, and Ultimate Design models. All new phones are powered by Kirin chips: the Mate 90 uses the Kirin 9030, the Pro uses the Kirin 9035, and the Pro Max uses the Kirin 9050 Pro. For Pro and above models, the lens cover uses the industry’s first optical Kunlun Glass, with drop resistance 20 times greater than the previous generation, and combines an integrated Xuanwu body with third-generation Kunlun Glass for all-around protection. The series debuts multi-user mutual-assistance communication sharing, supporting up to four phones helping one another at once and aggregating the network capabilities of multiple Mate devices. On imaging, all models feature the third-generation Red Maple true-color camera with infrared optical coating, adding Fireworks Portrait, Soft Light Portrait, Starburst effects, and more. It debuts the industry’s first sensor-equipped 10x continuous optical zoom module camera and adds a Professional Livestream Mode with a self-developed switcher. The system runs HarmonyOS 7, with more than 450,000 HarmonyOS apps and services and 35,000 HarmonyOS games; the number of devices running the single HarmonyOS framework has surpassed 90 million. On pricing, the Mate 90 starts at 5,999 yuan, the Pro at 6,999 yuan, the Pro Max at 9,499 yuan, the Pro Max Collector’s Edition at 10,999 yuan, and the RS Ultimate Design at 12,999 yuan. Yu Chengdong said that with memory prices soaring today, the pricing is still very sincere. @@@ 36Kr reports that the 2026 Apsara Conference is sending a signal of generational handover in Alibaba’s AI. At the main forum, Liu Dayiheng, born in 1993, spoke after Joseph Tsai and Eddie Wu. He is the head of the Token Foundry Qwen LLM project in Alibaba’s ATH business group and the newly confirmed leader of the Tongyi Qianwen large model. Joseph Tsai said AI must shift from technological imagination to large-scale real-world deployment; Eddie Wu said total machine thinking is currently less than 3% of humans and will reach more than 1,000 times humans in the future, and mentioned that global data centers will exceed 20 GW by 2032. Liu completed his bachelor’s, master’s, and doctoral degrees at Sichuan University. After joining Qwen in 2021, he led pretraining and participated in the full Qwen1 to Qwen3.5 series and the open-sourcing of more than 300 models, with over 36,000 Google Scholar citations. At the Qianwen Office forum, Chen Yusen, born in 1992, made his first official appearance as the youngest business unit CEO in Alibaba’s history. He founded Chaitin Tech and sold it to Alibaba Cloud, returned to Alibaba Cloud in 2025 to incubate MuleRun, and became DingTalk CEO in June this year; in July, Alibaba integrated QoderWork, Wukong, and MuleRun to form the Qianwen Office business unit, which he heads. @@@ 36Kr, citing a September 30 report from Zhidx, says that on September 29, Beijing AI Infra company DataCanvas submitted an IPO application to the Hong Kong Stock Exchange, seeking a main-board listing, with China Galaxy International Securities as sole sponsor. Founded in 2013, the company transformed from a data science platform into an AI computing infrastructure provider and is advancing a Token Factory strategy. By 2025 revenue, it ranks ninth among China’s intelligent computing service providers and is the only “new intelligent computing cloud” in the top ten. Revenue from 2023 to 2025 was 106 million, 396 million, and 1.098 billion yuan, respectively, growing 933% over three years and up 177.7% year over year in 2025; revenue in the first half of 2026 was 588 million yuan, up 92% year over year. Net losses from 2023 to 2025 were 394 million, 458 million, and 336 million yuan, respectively. In the first half of 2026, net loss was 1.095 billion yuan, of which 1.068 billion came from non-operating factors, and adjusted net loss narrowed to 14.88 million yuan. As of the latest practicable date, controllable computing power exceeded 29,000 PFLOPS, serving more than 150 enterprise customers; in 2025, top five customers contributed 88.0% of revenue. The company has completed 14 financing rounds, with investors including Beijing Guoguan, Zhongguancun Capital, and others. @@@ 36Kr reports that people familiar with the matter say Tencent Yuanbao will launch a personal agent, though product form and launch timing are undecided. The personal agent runs on a user-exclusive independent cloud virtual machine, can work 24/7, and supports long-term memory, cross-session execution, proactive responses, and long-horizon tasks. Yuanbao’s specific plan is undecided; it may undergo a complete overhaul and abandon its chatbot identity, or be added as a new module to the existing product. The move comes as Yuanbao gradually yields the spotlight inside Tencent: the group is focusing on office Agent platform WorkBuddy, and earnings reports praise it more while devoting less to Yuanbao. During the Spring Festival red-envelope battle, Yuanbao’s DAU was once claimed to exceed 50 million, but it fell rapidly after subsidies disappeared; QuestMobile data showed it was below 10 million by the end of March. But Yuanbao still faces the challenge of reshaping user perception, as well as funding and computing power pressure. Tencent is preparing multiple moves: the rumored new app Handy Bot, whose WeChat service account has been canceled; the LightVela from Tencent’s lightweight cloud team has entered public beta, offering cloud Agent hosting, long-term memory, and a dedicated multi-cloud computer, with new users getting a one-month free trial and 4,500 AI credits. ByteDance’s Spell project has accelerated due to Muse’s explosive popularity and will soon launch a standalone app, Xiaodou. @@@ 36Kr reports that Mech-Mind founder Shao Tianlan questioned on WeChat Moments on September 9 that some high-valuation embodied intelligence companies use data collection centers and related-party transactions with local governments, investors, and suppliers to create fake, unsustainable revenue through a model in which “local state capital buys robot bodies and companies buy back data” in order to sprint toward an IPO. As of August 2026, China had built 106 centralized embodied data collection centers, 84 of which are operational, yet a stable commercial loop has not yet formed. Lingyu Intelligence CEO Jin Ge said almost all data collection centers use this model; his company ships nearly 100 units per month and more than 300 cumulatively, while the Yibin training center purchased 30 units, less than 10% of shipments. Data collection center customers account for nearly 50% of units, and together with real-world customers contribute 70% to 80% of order revenue. Songyan Dynamics founder Jiang Zheyuan said the company has no orders from data collection this year and believes that requiring data buybacks is not real. Shuimu Capital’s Tang Jincao said institutions have slowed their pace, conducting due diligence with cross-validation of utilization rates, payers, and related-party relationships, and at worst excluding such revenue from core valuation. @@@ According to 36Kr citing Machine Heart, Google announced its new-generation frontier model Gemini 4 Argon, positioned for complex, long-horizon workflows and covering software engineering, legal and financial enterprise knowledge work, and cybersecurity defense. The model is being released in phases, with the first batch provided through the Fairwind Program to trusted cybersecurity defenders. Initial API pricing is $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower. Its single-output cap has been raised from 64,000 to 1 million tokens, far above the roughly 128,000-token cap of most frontier models. Google says Argon scored 77.9% on the DeepSWE v1.1 software engineering evaluation, ranking first on the Vals Index; it scored 51.3% on AutomationBench and 91.7% on LVBench long-video understanding. Thousands of internal employees have already used it for code debugging, algorithm design, and large codebase migration, and in quantum computing tests it improved benchmark results by 40% in minutes; data center memory optimization has freed more than 300 TiB. On cybersecurity, Wiz used Argon through Scan for Good and once found a severe vulnerability affecting medical software used by hospitals worldwide, tying for first with a score of 68% on CWE-bench v1. Google has also deployed mechanisms such as abuse limits, red-team testing, and task-abort mechanisms. But Bloomberg reports that Google internally still has doubts about Gemini 4’s performance in coding and other key tasks. After completing early testing, Argon will gradually open to developers, enterprises, and consumers; the first public users include paid API customers and Google AI Ultra subscribers. @@@ According to 36Kr, OpenAI safety veteran David Robinson resigned and wrote in The Atlantic warning that “the era of trial and error is over,” saying OpenAI’s culture has collapsed. He worked there for three and a half years, led the drafting of the Preparedness Framework, and wrote safety reports for 12 frontier model releases, but said he never met a colleague with aviation, nuclear power, or financial safety experience. He disclose
