Alibaba2026-09-22 07:22:21Alibaba says Qwen4 is in training as later model roadmap stretches to 10 trillion parametersAlibaba used its annual Apsara Conference in Hangzhou to lay out a broad AI and infrastructure roadmap, centered on the next generation of its Qwen models. The company said Qwen4 is now in training, while later releases in the line, including Qwen4.5 and Qwen5, are expected to scale to between 5 trillion and 10 trillion parameters. Alibaba also shared updates on recursive self-improvement, saying Qwen3.8-Max completed 33 iterations over more than a month of fully automated operation and improved its Artificial Analysis score from 40 to 45. In a chip-design experiment, the model reportedly ran self-improvement for more than 60 hours, called EDA tools over 10,000 times, produced a production-grade chip bus module, and cut chip area by 42% without changing performance. On the hardware side, Alibaba affiliate T-Head unveiled the Zhenwu V900 AI processor, which it said delivers three times the performance of the Zhenwu M890 launched in May. Alibaba also introduced a new audio model, previewed an image model due later this year, released a cross-app agent platform for smartphone makers, and set a 2032 target of more than 20GW in global data center capacity for Alibaba Cloud.680
Alibaba2026-09-22 02:21:40Alibaba plans bigger AI infrastructure push as T-Head unveils Zhenwu V900Alibaba is stepping up its artificial intelligence buildout, with CEO Eddie Wu saying at the Apsara Conference that customer demand for AI remains strong and the company will commit fully to AI infrastructure. Wu said Alibaba aims for the global data centers operated by Alibaba Cloud to exceed 20 gigawatts by 2032. At the same event, T-Head introduced its Zhenwu V900, described as its strongest domestically developed AI chip. The company said the chip delivers three times the computing power of the Zhenwu M890, and projected a sharp increase in annual shipments. The announcements point to Alibaba’s focus on expanding both cloud capacity and in-house computing hardware as AI demand continues to rise.530
CXMT2026-08-31 15:15:05CXMT starts small-batch HBM3E production, with volume expansion planned for 2027Chinese memory maker ChangXin Memory Technologies, or CXMT, has started small-batch production of HBM3E and plans to expand output in 2027, according to The Information, which cited two people familiar with the matter. The report said domestic chip teams including Cambricon and Alibaba’s T-Head have already been testing CXMT’s HBM products, and if those efforts go smoothly, the memory could appear in products as early as next year. HBM, or high-bandwidth memory, sits alongside AI chips and feeds them data at high speed, making it a critical component for AI computing. The report described HBM2 and later generations as a longstanding weakness for China’s domestic AI chip supply chain. In product-generation terms, CXMT is one step behind SK Hynix, Samsung and Micron, which are already mass-producing HBM4. Still, the report said small-batch production is not the same as mature mass production. CXMT’s HBM3E currently has lower yields, and its speed and reliability trail the three established suppliers. SK Hynix began mass production of HBM3E in 2024, and the three leading makers have already moved into HBM4 and started sampling HBM4E.1190
