Meituan's LongCat-2.5-Preview goes live with 1.6T parameters Meituan's LongCat API platform launched LongCat-2.5-Preview on September 25. It's a mixture-of-experts model with about 1.6 trillion total parameters, activating roughly 48 billion per inference. The model adds image understanding, claims strong coding performance, and supports a 1 million token context window natively. The pitch is long-horizon tasks: ultra-long documents, codebases, logs, multi-turn complex work. Meituan says it integrates with Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code. Our read: Meituan wants developer mindshare, not just internal logistics AI. The 1M context and coding focus target agentic workflows where Meituan's local commerce data could eventually differentiate. But this is a preview release with no benchmark numbers yet, so actual capability is unproven. We'd argue the more interesting question is whether LongCat-2.5 can turn that parameter count into reliable long-task performance. Only 48B parameters activate per token, which leaves a lot of dormant capacity to manage. $MPNGY

Alibaba names Liu Da Yiheng to lead Qwen Alibaba has put Liu Da Yiheng, a former Huawei "genius youth" researcher, in charge of its Qwen large-model team. He replaces Lin Junyang. $BABA

Alibaba, ByteDance and DeepSeek All Chase 10-Trillion-Parameter Models At Alibaba's Yunqi conference on September 22, CEO Eddie Wu said the company is preparing to train a new model with 5 trillion to 10 trillion parameters. Alibaba's Liu Dayiheng put Qwen 4.5 and Qwen 5 on the same 5T-10T trajectory. The Information reported the same day that DeepSeek is training a roughly 2-trillion-parameter model, with Liang Wenfeng telling investors the next step is 8 trillion. In August, the Financial Times reported ByteDance is training a model of up to 10 trillion parameters. For scale: DeepSeek-R1, the model that set off the 2025 wave, had 671 billion parameters, less than one-fifteenth of 10 trillion. The bill is the story. Our read: the parameter number is mostly a claim on the next generation's starting position, not a promise about the current one. That is why the same labs can push Flash-style cheap models and 10T flagships at once. What that architecture does not decouple is the data bill. We're not convinced the 10T framing survives contact with deployment. The race is real. h/t @TMTPostGlobal

Unitree founder: robot bottleneck is millimeter-level work error At the Global Digital Trade Expo in Hangzhou on September 24, Unitree founder Wang Xingxing said the core bottleneck for humanoid robots is whether they can resolve work errors of a few millimeters. His benchmark for the industry's "ChatGPT moment": a robot that can complete 80% of tasks in 80% of unfamiliar scenarios via voice and embodied capabilities. Wang said robots can already follow spoken instructions to perform specific tasks. But AI model input and output don't match the physical world precisely enough, leaving millimeter-level errors. "Whoever solves this problem solves the robot problem completely," he said. Our read: the 80/80 threshold is a useful framing, but it masks the harder question of whether the error is a software or hardware problem. Millimeter-level precision in unstructured environments is as much about actuators and sensing as it is about model generalization. Unitree's own admission that new tasks require retraining suggests the bottleneck is precision and adaptability. Wang made similar comments at WRC 2026 in August, saying efficiency and capability are still too low for large-scale deployment in factories and homes. He wants more generalized technology and more precise capabilities before pushing into specific scenarios.

Huawei's Ascend 950 supernode hits commercial scale Huawei launched the Ascend 950 supernode at its Connect 2026 conference. It calls the product the industry's largest commercially deployed supernode. The liquid-cooled Atlas 950 SuperPoD holds 1,024 NPUs with unified memory addressing and 2.1x higher interconnect bandwidth. Huawei claims 1.5-1.7x training efficiency gains. It also claims sub-10ms inference latency for trillion-parameter MoE models. A companion air-cooled Atlas 850E fits standard enterprise racks. Companies skip the cost of liquid-cooling retrofits. Our read: Huawei is shifting from selling chips to selling compute as a service. This launch extends that. The hard part is no longer the NPU count. It is the software and network reliability that make 1,024 cards behave like one machine. The orthogonal architecture removes 50 km of copper per node. Liquid-cooled optical modules cut power and latency. That is what separates a paper supernode from one enterprises can actually deploy. H/t @Chinazhidx

Unitree tops both quadruped robot rankings, but the No. 2 is disputed On September 21, Counterpoint and IDC each published a global quadruped robot market report. Unitree is first in both. After that, the lists stop agreeing. Counterpoint ranks by shipments and puts Zhiyuan Kutu second. IDC ranks by hardware revenue and puts Deep Robotics second. In industry applications, Unitree drops to fourth by shipments, and the two firms again name different leaders. The totals are just as far apart. Counterpoint counts nearly 35,000 units shipped in H1 2026, with 40% going to industry use, and 46,000 units for all of 2025, with industry revenue near $700 million. IDC counts over 49,000 units in H1 2026, up 92.5% year-on-year, and about 58,000 units in 2025, with a market size over $490 million. The 12,000 to 14,000 unit gap is more than half of Unitree's 23,000 quadruped sales in 2025. The gap comes from what each firm counts. Counterpoint includes sensors, scheduling software, and maintenance services in market size. IDC counts only hardware. Counterpoint also excludes small entry-level companion robots, while IDC covers consumer products. Unitree's average quadruped price in 2025 was RMB 30,300 (~$4.2K), so its revenue lead over Deep Robotics narrows to 4.6 percentage points. Deep Robotics sells its industrial Jueying X series for RMB 287,500 (~$39.7K) per unit. We'd argue the rankings say less about who is winning than about how fragmented the market still is. No player has a product that spans consumer, education, and industrial segments. Unitree's flip from shipment leader to narrow revenue leader shows price strategy, not volume, decides the top spot. The more interesting race is whether quadruped robots can lock in industrial inspection and security budgets before humanoid reliability and cost improve enough to compete for the same money. h/t @TMTPostGlobal

DeepSeek's revenue run rate hits $1 billion as new funding round nears DeepSeek's annualized revenue run rate has reached $1 billion, founder Liang Wenfeng told investors recently, up from under $500 million a few months ago. The jump follows price hikes on its flagship V4-Pro model in August and sustained demand, even as its lighter V4.1-Flash got cheaper. The company is also finalizing a new RMB 50 billion ($7 billion) round at a RMB 500 billion ($70 billion) valuation, targeting an October close, with a STAR Market IPO in parallel. That follows a June round of similar size at a RMB 400 billion (~$55.2B) valuation. Our read: the revenue acceleration is real, but the more telling number is that over 70% of compute still goes to training, not serving paying customers. DeepSeek is deliberately leaving inference capacity constrained while it chases a 2 trillion parameter model. The bet is that frontier capability, not API volume, is what justifies a $70 billion valuation. We're not convinced that math works if inference demand keeps growing faster than the chips to serve it. h/t @Chinazhidx

CATL's moat in EV batteries is under siege from BYD and emerging cell makers. The company defends share through scale and R&D, but rivals are closing the gap. Our read: the real test is whether CATL can keep pricing power as capacity floods the market.

Overseas demand to force China auto consolidation LatePost argues that 10 million incremental overseas vehicles will trigger a wave of consolidation in China's auto industry. Global EV demand growth becomes the catalyst reshaping domestic automaker competition and M&A. H/t @latepostnews