|
|
|
|
Bottom Line Up Front Alibaba Cloud unveiled the Zhenwu M890 AI chip alongside the Qwen3.7-Max model and a 128-card super-node server at its annual summit, with T-Head reporting cumulative chip shipments of 560,000 units across 400 enterprise customers and a roadmap targeting the V900 in Q3 2027 and J900 in Q3 2028. The National Data Administration issued a six-initiative framework for high-quality AI training datasets at a May 19 Guangzhou meeting, covering foundation expansion, annotation, AI-Ready quality standards, application, lifecycle management, and a token-based data value system, framed as implementing CPC Central Committee and State Council AI decisions. NDRC Chair Zheng Shanjie convened a May 20 private economy roundtable seating Zhipu AI (Zhipu Huazhang) alongside Delong Steel, Hengrui Medicine, Dayu Water Saving, and the Yi-Xin-Ou rail operator, where participants raised intelligent compute coordination as a policy ask and Zheng identified tech self-reliance, supply chain self-control, anti-involution measures, and effective investment as NDRC priorities under the 15th Five-Year Plan Outline. |
| Alibaba Cloud rolls out Zhenwu M890 AI chip and Qwen3.7-Max in full-stack agent upgrade Alibaba Cloud announced at its annual summit a "chip-cloud-model-inference" agent upgrade built around the new Zhenwu M890 AI chip from semiconductor subsidiary T-Head, a 128-card super-node server, and the Qwen3.7-Max flagship model now live on the Bailian platform, Wallstreetcn reported. The Zhenwu M890 carries 144GB of memory, 800GB/s inter-chip bandwidth, and native support spanning FP32 to FP4 precision, with the matched ICN Switch 1.0 linking 128 chips into a single compute unit at 100-nanosecond class communication latency. T-Head Vice President Gao Hui said cumulative chip shipments have reached 560,000 units across more than 400 enterprise customers in 20 industries, up from a previously disclosed 450,000. T-Head's roadmap calls for the V900 in Q3 2027 at roughly three times M890 performance and the J900 in Q3 2028. |
| National Data Administration codifies six-initiative framework for high-quality AI training data The National Data Administration convened a May 19 on-site meeting in Guangzhou to roll out a six-initiative architecture for high-quality dataset development, Yicai reported citing the agency. The named initiatives are foundation expansion, annotation breakthrough, quality improvement to "AI-Ready" requirements, application empowerment, full lifecycle management with a "physically distributed, logically centralized" resource catalog, and value release through a "token-based data value system" and a "pay for high-quality data" market norm. The framework is explicitly framed as implementing CPC Central Committee and State Council decisions on AI development, and follows the May 18 launch of the Greater Bay Area Data Special Zone Innovation Cooperation Center. |
| NDRC seats Zhipu AI at private economy roundtable, names intelligent compute coordination as a 15th Five-Year Plan priority NDRC Chair Zheng Shanjie convened a private firms roundtable on May 20, the first anniversary of the Private Economy Promotion Law, with Zhipu AI (legal name Zhipu Huazhang) taking the AI seat alongside Delong Steel, Hengrui Medicine, Dayu Water Saving, and the Yi-Xin-Ou China-Europe Railway operator, Wallstreetcn reported. Participants raised "strengthening intelligent compute coordination" and soft-hardware integration in national major projects as policy asks. Zheng named "tech self-reliance, supply-chain self-control, deep rectification of involution-style competition, and effective investment" as next-step NDRC priorities under the 15th Five-Year Plan Outline. |
| DeepSeek confirms Harness team to build desktop coding agent against Anthropic's Claude Code DeepSeek senior researcher Chen Deli confirmed via social media that the lab is forming a dedicated Harness team to build a coding agent product internally positioned against Anthropic's Claude Code, CnBeta reported citing Jiazi Guangnian. The job spec frames the work as DeepSeek's "desktop Agent product" and codifies the internal formula "Model + Harness = Agent," with Harness defined to cover everything beyond the model itself. DeepSeek has opened two Beijing roles, Harness product manager and Harness research and development engineer, with its office in Haidian's Rongke Information Center near Peking and Tsinghua universities. |
|
| Qwen3.5-LiveTranslate compresses real-time speech translation to 2.8-second per-character latency Alibaba's Tongyi Lab released Qwen3.5-LiveTranslate-Flash, a real-time speech translation model supporting 60 languages on audio input and text output and 29 languages on audio output, ITHome reported citing Tongyi Lab. The model claims 2.8-second end-to-end per-character latency through a new "Readable Unit" streaming output technique, and adds dynamic cross-lingual voice cloning that preserves speaker timbre across translated outputs. Built-in hot-word capability targets proper nouns and industry terminology, with the release rolled out alongside the Alibaba Cloud summit cadence on May 20. |
| openJiuwen open-sources JiuwenSwarm, advancing Coordination Engineering as a multi-agent approach The Huawei-supported openJiuwen open-source community open-sourced JiuwenSwarm, a swarm-agent framework for multi-agent collaboration, InfoQ China reported. JiuwenSwarm targets role assignment, task decomposition, experience accumulation, and failure feedback loops, and is built on top of the openJiuwen Harness substrate. openJiuwen frames the lineage as Prompt Engineering, then Context Engineering, then Harness Engineering, and now Coordination Engineering. |
|
| Approaching.AI closes Pre-A round on a "Token as a Service" inference infrastructure pitch Approaching.AI announced a Pre-A round of several hundred million yuan led by Star Connect Capital and Huakong Technology, with existing investor Hillhouse-affiliated GL Ventures continuing to back the company, Wallstreetcn reported. The company describes its ATaaS platform as handling near-trillion daily token invocations and pitches "Token as a Service" as a deliberate alternative to model-as-a-service for enterprise AI inference. Approaching.AI's technical stack centers on a "Liuhe" heterogeneous inference architecture and a "Mooncake" storage-for-compute substitution that reconstructs KV-Cache to reduce GPU dependency. CEO Ai Zhiyuan is a Tsinghua PhD with a parallel computing background who previously led research and development at Sangfor. |
|
| MFA: China and U.S. will "lock in" preliminary trade results "as soon as possible" MFA spokesperson Guo Jiakun told a May 20 briefing that China's competent authorities have already introduced preliminary results of U.S.-China economic and trade consultations and that both sides will move to lock in those results "as soon as possible," Yicai reported citing CCTV News. Guo said Xi and Trump set the direction during the Beijing summit and that implementation will be a joint responsibility. The MFA framed the lock-in as injecting "more certainty and stability" into U.S.-China trade cooperation and the world economy. |
| Huawei pairs IDC two-year tablet share lead with a "processor-hardware-OS-cloud" self-reliance pitch Huawei tablet and PC line president Zhu Dongdong said HarmonyOS Office is leaning on "soft-hardware-chip-cloud collaboration" to compete on enterprise productivity, ITHome reported from the May 20 HarmonyOS Office briefing. Huawei cited IDC data ranking its tablets number one in China shipments for two consecutive years across 2024 and 2025, and said 2025 Huawei Qingyun enterprise service customers exceeded 100,000. Huawei said its self-reliance scope spans processors, hardware, operating system, and cloud. |
|
|
|