Rotating chairman David Wang told Huawei Connect 2026 in Shanghai that the Ascend 960DT accelerator will ship in the first quarter of 2027, three quarters earlier than the company's previously published schedule, per Pandaily. A second variant follows in the third quarter. Wang also introduced a supernode that ties about 4,096 accelerator cards into what behaves as a single machine while holding power consumption down at that scale. Huawei said more than 1,000 of its supernode systems have already shipped to more than 370 customers. Wang put million card clusters on the longer term roadmap and said Huawei has developed 11 of its own chips for linking those systems together.
Read at Pandaily ↗
People's Daily (state media) ran a commentary under the pen name Zhong Sheng arguing that Washington has recast a routine engineering practice as a national security threat to protect an industrial monopoly, per the newsletter Sinocism. At issue is distillation, the training of one model on another model's outputs, which U.S. security agencies said this month Chinese firms had run at industrial scale against American systems. The commentary said American companies use the same method and call it standard practice, and described the asymmetry as "naked double standards." It faulted the broad geographic restrictions U.S. firms write into user agreements. The commentary closed by repeating that Beijing will take countermeasures if Washington acts against Chinese AI companies over distillation.
Read at Sinocism ↗
A Chinese built forecasting model can now produce 10 day global forecasts at 5 kilometer (3.1 mile) resolution, against roughly 5 to 10 kilometers for most major systems in routine use worldwide, the China Meteorological Administration said, per the South China Morning Post. The model runs on Sugon 8000, a 100,000 card system assembled entirely from domestically made AI accelerators that Sugon showed at the World Artificial Intelligence Conference in Shanghai in July. One 10 day run at that resolution finishes inside an hour, which the research team said meets international operational benchmarks. Researchers have also completed a global simulation test at 3 kilometer resolution.
Read at South China Morning Post ↗ • Read at Tencent News ↗
Xiaomi has opened a public dashboard streaming the reinforcement learning phase (training a model by scoring its own attempts) of its next model, MiMo-V2.6, showing money spent, steps completed, reward curves and which nodes have suffered chip failures, per QbitAI. The two variants had together cost more than $1.08 million in the 36 hours after the Pro model began training, averaging about $30,000 an hour. Luo Fuli, who leads the work, said the roughly six months since Xiaomi open sourced its previous model went into one question, how far reinforcement learning can be scaled. The three levers she named are training compute, the number of task environments a model works in, and the compute spent grading its attempts. Xiaomi's own offline scores on a software engineering benchmark put its two variants at 62.24 and 60.77, against 74.2 for DeepSeek's Flash model.
Read at QbitAI ↗