China AI updates span long-context models, liquid cooling and advanced packaging
China’s AI compute news centered on foundation-model releases, data-center cooling and power-grid coordination, and domestic packaging progress.
• Xiaomi MiMo head Luo Fuli said MiMo-V3 will adopt a new architecture, while its HySparse 2 component was released with 5.02x lower prefill FLOPs and a 4.5x smaller KV cache at 1 million tokens.
• Alibaba Qwen released the Qwen-Audio-3.1 speech model series, adding new audio creation and audio understanding models and cutting prices across the Qwen-Audio product line.
• iFlytek released Spark-ASR-2.0, a speech recognition model aimed at mixed Chinese-English, dialects, specialized terminology and difficult acoustic scenarios.
• Ant BaiLing open-sourced the Ming-Image-0.1-Design series, including two 6B-parameter models for visual design generation and editable layer decomposition.
• China’s cyberspace regulators announced filings for three on-device mobile generative AI services: Honor YOYO Claw, Xiaomi miclaw and StepFun Terminal AI.
• AAC Technologies and Intel formed a liquid-cooling ecosystem partnership covering high-density server and rack-level technologies including cold plates, memory and SSD cooling, quick connectors, manifolds and CDUs.
• China’s latest information and communications industry plan targets 9,800 hundred-quadrillion FLOPS of intelligent computing power by 2030, alongside 95% 5G user penetration and 320 million broadband users above gigabit speeds.
• BOE said its glass-based package substrate pilot line has completed end-to-end operation, with 24-layer products larger than 100 x 100 mm delivered to customers for validation and testing with real CPU and GPU chips in servers.