Latest headlines from Pandaily, covering world, with clear attribution and direct links to the original publisher.
XPeng is pushing its self-developed Turing AI silicon beyond intelligent EVs into humanoid robots as part of a Physical AI stack, with IRON configurations cited at about three on-device Turing chips totaling roughly 2,250 TOPS.
Boyagongdao displayed its BG-5 Golden Dragon Fish biomimetic AUV at CIFTIS — about 3 kg and 640 mm long, with a soft multi-joint tail, 0.6 m/s top speed and a 5-meter depth rating for propeller-free shallow inspection.
Blueworm Embodied introduced Mantis Standard, a modular wheeled humanoid starting at RMB 9,800, with 22 degrees of freedom, a 7 kg single-arm payload, an omnidirectional 3 m/s chassis and teleoperation latency under 10 ms.
DeepSeek released DeepSeek Harness, an MIT-licensed agent runtime in developer preview where models, tools, sessions, sandboxes and UI are Cordis plugins, with npm packages advancing through the v0.1.5-rc line.
At IAA Transportation 2026, Pony.ai and GAC Commercial Vehicle debuted a Gen-4 Level-4 electric heavy truck on the GAC T9 platform, aiming for volume production later in 2026 with a lower-cost automotive-grade ADK.
Axera launched its 5nm M9 ADAS SoC series — flagship M97 at about 720 TOPS and mainstream M95 at about 360 TOPS — with self-developed NPU, ISP and CVA blocks and dual-chip redundancy near 1,440 TOPS.
Tsinghua AIR, with Domain Transform, introduced Zeva — frozen-weight in-context causal memory that raised cumulative embodied manipulation success from about 26% to 73% on Cosmos3-backed ChemLab-Evo tasks.
At the China Computing Power Conference, Tsingway and the Beijing Academy of Artificial Intelligence open-sourced Open3D-PIMC, a programming model and framework for 3D compute chips inside the FlagOS stack.
Unitree refreshed its G1 humanoid as G1+, adding neck degrees of freedom, stronger cooler motors, dual vision plus tactile and six-mic audio, and an external power port, with the standard model listed around RMB 95,000 tax-included.
Unisound unveiled U2-Flash, a ~266B sparse MoE with ~10B active parameters, pairing post-training recursive self-improvement and latent reasoning for agent and coding workloads, now live on its MaaS platform.
Moonshot AI says Kimi K2.8 Preview is fully live for Kimi Code and Work users, with coding and agent gains near flagship K3, adjustable thinking, 1M context on all tiers, and the unchanged API ID kimi-for-coding.
Shanghai Artificial Intelligence Laboratory released Intern physical world model W0 with native force-tactile sensing and duplex collaboration, linking it to Intern InkStone and science model S2 for wet- and dry-lab closed loops such as lipid nanoparticle synthesis.
Shengshu Technology (GensPI), with Tsinghua-affiliated researchers, unveiled Motus2 — a shared-parameter world-action, action-conditioned world and value model loop for dexterous manipulation, lifting phone-placement success from about 65% to 75% with ~130k hours of egocentric data.
DeepSeek is gray-testing in-app AI voice chat: a speaker control in the top-right corner and four selectable TTS voices — Beike, Bailang, Haixing and Anchao — as a product feature distinct from Flash GA and the Pro-to-Flash API cutover.
At the 2026 China Computing Power Conference, China Mobile Cloud and partners unveiled a domestic GPU plus neuromorphic mixed-inference system for large models, citing roughly 2× output and energy gains and over 40% lower opex on DeepSeek V4 Flash.
