Skip to content
Source · Daily Brief

AI Daily Brief — 27 July 2025

WAIC day 2 — arguably the loudest news day in Chinese AI since DeepSeek-R1. Tencent shipped, SenseTime shipped, Huawei detailed, and the domestic stack visibly circled the wagons under US export curbs with formal Model-Chip alliances.

Top stories

  • Tencent Hunyuan 3D World Model 1.0 stays the lead. Industry’s first fully open-source AI model generating explorable, interactive 360-degree 3D scenes from text or image prompts. Weights, code and docs released; outputs exportable as 3D meshes for Unreal/Unity/Blender. via GitHub
  • SenseTime launches SenseNova V6.5 at WAIC Large Model Forum. “Rixin” multimodal foundation model with interleaved text-image chain-of-thought (claimed industry-first commercial-grade). Reasoning on par with Gemini 2.5 Pro and Claude 4 Sonnet, with ~3x cost-performance gain via early encoder-level fusion (visual encoder shrunk to 1B params). via SCMP
  • Tencent launches Tairos embodied-AI platform. China’s first modular embodied-intelligence software platform. Bundles foundation models (task planning, multimodal perception, perception-action integration) plus cloud services for simulation/training/data. Plug-and-play infrastructure for robot makers rather than competing hardware. via KrASIA
  • Huawei CloudMatrix 384 public debut detailed. 384 Ascend 910C chips linked via optical interconnects, claimed 300 PFLOPS dense BF16 (roughly 2x NVIDIA GB200 NVL72) with 3.6x memory capacity and 2.1x bandwidth. Marketed as China’s domestic alternative to NVIDIA training clusters amid US export curbs. via IEEE ComSoc
  • Chinese AI firms form Model-Chip Ecosystem alliances. StepFun convened the “Model-Chip Ecosystem Innovation Alliance” pairing Chinese LLM developers with domestic chipmakers Huawei, Biren, Moore Threads and Enflame. A second “Shanghai General Chamber of Commerce AI Committee” — including StepFun, SenseTime, MiniMax, MetaX and Iluvatar CoreX — also established. via Yahoo/Reuters

Who shipped

SenseTime shipped V6.5. Tencent shipped Tairos. Huawei shipped CloudMatrix detail. The Chinese AI stack shipped two alliances.

Open-source pulse

Alibaba unveiled Quark AI Glasses (Qwen-powered, hands-free calling, real-time translation, end-2025 China launch) plus a Qwen-powered automotive LMM with Qualcomm and Banma Network. Baidu showcased ERNIE-powered NOVA digital-human platform — virtual livestreamers from ~10 minutes of sample footage.

Money, infra & hardware

MiniMax CEO Junjie Yan delivered “Everyone’s AI” WAIC keynote arguing for democratised AI access — citing M1 reasoning model trained on just 512 GPUs over three weeks at total cost $534,700. via MiniMax

Quiet corners

Geoffrey Hinton’s “tiger” warning dominated WAIC Sunday news cycle — goal-pursuing AI agents will seek self-preservation and control; “train it to be tame or be prepared to kill it”. Building on Premier Li Qiang’s 13-point Global AI Governance Action Plan from July 26, Beijing formally proposed creating the World AI Cooperation Organization (WAICO).

By the numbers

  • 1B — SenseNova V6.5 visual encoder param count (shrunk for efficiency)
  • ~3x — V6.5 cost-performance gain
  • 384 / 300 PFLOPS — Huawei CloudMatrix Ascend chips / dense BF16
  • ~10 min — Baidu NOVA cloning footage needed
  • 512 GPUs / 3 weeks / $534,700 — MiniMax M1 training profile

Compiled by AI Feed’s editor from verified web sources for 27 July 2025.