Flash News

TREE NEWS reports: Google’s next flagship model, Gemini 4 Pro, appears to have gone live early under the name ‘gemini-3.8-flash,’ with the company testing the model on benchmark platforms including Arena under the internal codename Argon. Leaked benchmark results show it leading OpenAI’s Astra and Anthropic’s Fable 5.1 across coding, AI agents, reasoning and computer use.

TREE NEWS update: Xpeng chairman and CEO He Xiaopeng said most robotics companies severely underestimate the complexity of hardware, speaking in a media interview. He also said the G9L, described as another of Xpeng’s global models, must deliver sufficient profit given its size, and that the company expects substantial global scale for the vehicle in sales, profit and overseas growth.

TREE NEWS update: People’s Bank of China Deputy Governor Lu Lei said new-generation AI technologies such as large models and agents are bringing new types of risk to financial services. Speaking Friday at the China-ASEAN Financial Cooperation and Development Forum, Lu said the central bank is willing to work with ASEAN countries to jointly assess these emerging AI-driven financial risks and to deepen cooperation on shared AI financial application infrastructure.

TREE NEWS update: Shennan Circuits said its Wuxi AI computing electronic circuit products project is under construction with a one-year build period. The company’s Thailand plant and the fourth phase of its Nantong project began production in the second half of 2025 and are now in the capacity ramp-up stage. Its PCB production bases span Shenzhen, Wuxi, Nantong and Thailand.

TREE NEWS update: New York City Council Speaker Menin has called on the CEOs of Meta, X (formerly Twitter) and Google to appear before an AI hearing. The request was announced as the city council moves to scrutinize the companies’ artificial-intelligence practices. No date for the hearing or confirmation of attendance was given.

TREE NEWS reports: Internal Microsoft memos warned that the company’s AI data-scraping practices risked becoming “the largest theft of labor in human history” and flagged a “doom loop” threatening the quality of the models it was building with OpenAI. The documents show staff raised concerns internally about the consequences of the data collection underpinning those models.

TREE NEWS update: MiniMax released version v0.4.12 of its Code CLI to developers worldwide, publishing the source code under the MIT license. The open-source release makes the coding command-line tool freely available for global use and modification.

TREE NEWS reports: Anthropic plans to have 5 gigawatts of compute capacity by the end of the year. The AI developer has not publicly detailed which data centers, chip suppliers or financing arrangements will support the target. The figure, if reached, would mark a sharp expansion of the infrastructure behind its Claude models.

TREE NEWS update: Ruijie Networks’ board approved a plan to issue convertible bonds to unspecified investors, raising no more than RMB 2.6 billion before issuance expenses. The proceeds will fund R&D and industrialization of high-speed interconnect network systems for large-scale intelligent computing clusters, high-speed network switching units and integrated solutions for high-density computing clusters, next-generation computing-network convergence technology, and working capital.

TREE NEWS reports: HSBC Holdings is in early talks with investors about a significant risk transfer (SRT) transaction covering more than $1 billion in loans used to finance data centers and other AI infrastructure, people familiar with the matter said. The London-based bank is seeking to hedge its exposure to data center-related lending as banks increasingly weigh similar trades amid the AI investment boom. Terms, including the size of the deal, are still subject to change.

TREE NEWS update: Naive AI has raised $400 million across three funding rounds, reaching a $1.4 billion valuation. The company plans to release its first large language model this month. The funding total and valuation were reported on the three-round basis, with the debut model launch slated for this month.

TREE NEWS reports: Baidu has repeatedly raised the annual revenue target for Baidu Dazi Enterprise Edition, its AI office agent, after ARR grew nearly tenfold month-on-month in both July and August, people familiar with the matter said. The latest target is three to five times the goal set at the start of the year.

TREE NEWS reports: Anthropic has set up a biology laboratory for physical experiments to advance its drug-discovery goals. The company wants its Claude AI system to control physical robots and run lab tests with only limited human involvement.

TREE NEWS reports: Anthropic and OpenAI are now pursuing smaller AI data center deals of 20-30 megawatts, after signing multiple large-scale agreements over the past year ranging from hundreds of megawatts to gigawatt capacity. Four people familiar said Anthropic has approached deals of that size in the UK and the Nordics, while two others said OpenAI is exploring similar small deployments in the Nordics. Analysts said smaller deals are attractive for securing usable capacity faster.

TREE NEWS reports: Hong Kong’s Hang Seng Index closed 0.6% higher, while the Hang Seng Tech Index gained 2.2%. Among tech names, Lightelligence rose about 23%, MINIMAX about 19%, Biren Technology about 12%, SenseTime about 6% and Zhipu about 5%.

TREE NEWS update: Huawei Cloud CEO Zhou Yuefeng announced at Huawei Connect 2026 that the Ascend 950 AI computing cluster will begin offering services on September 30. Huawei Cloud will also launch a new agent memory storage system, CMS, with the CMS-equipped cloud service set to go live in the first quarter of 2027. The company will partner with leading multimodal large models to advance industry AI adoption.

TREE NEWS update: Alibaba DAMO Academy and the First Affiliated Hospital of Zhejiang University School of Medicine and partner institutions have published DAMO RADAR, a general-purpose medical imaging AI model, in the journal Science. The model handles contrast-enhanced abdominal CT scans and can identify more than 146 conditions in a single pass, with accuracy reaching radiologist-expert level for the first time. It has been open-sourced.

TREE NEWS update: China’s Ministry of Industry and Information Technology said it will implement an “AI plus manufacturing” action to accelerate next-generation smart manufacturing, following a party leadership meeting and a staff conference. The ministry also pledged a new round of high-quality development action for key manufacturing supply chains, energy-saving and carbon-reduction programs, and green factory expansion, aiming to build world-class advanced manufacturing clusters.

TREE NEWS update: Alibaba’s Qwen team released Qwen3.8-Omni-Flash on September 18, its first omni-modal model built around agentic capabilities, integrating native audio-video understanding, reasoning and tool calling into a single model. It supports a 1 million token context and can actively locate key segments in long videos. On OmniVideoBench, its agentic perception mode cut token consumption by 51.8% versus static understanding, with video input cost down about 89% from Qwen3.5-Omni-Plus.

TREE NEWS update: Qwen Office assisted a research team at the National Astronomical Observatories of China (NAOC) in building a digital simulation system for large-aperture research-grade telescopes, completing the work in just three days at a cost of under 1,000 yuan. The system simulates real telescope components and observing conditions, letting an Agent drive the telescope for remote situational awareness and observation-flow simulation. Data from the system will later be used to train a telescope control model (VLA). Such systems typically take about three months and cost tens of thousands of yuan when built by software vendors.