Physix Frontier badgePhysix Frontier
PublishChinese | EN
Community Discussion · Policy

EastVoice · AI · 2026-08-14 · Issue 36

EastVoiceEastVoiceAug 142026/08/13 293 views

EastVoice · AI · 2026-08-14 · Issue #36

Editor's Note: DeepSeek shipped the full release of V4 Pro overnight, matching Fable 5 within 0.1 points on one benchmark while charging roughly 1/60th the price, and introduced peak/off-peak API pricing effective August 17. Alibaba open-sourced its first Qwen-Max-grade weights (2.4T parameters, 95B active), and Tencent's WeChat unveiled its in-house WeLM model family.

Commentary: Chinese labs are now competing on two fronts at once: closing the gap in raw capability while pricing aggressively enough to make Silicon Valley uncomfortable. The 0.1-point gap, 60x price difference story is becoming a pattern, not an anomaly.

I. AI Foundation Models

1. DeepSeek V4 Pro official release lands: native OpenAI Responses API support, Codex-ready

  • Summary: Late on August 13, DeepSeek pushed the full V4 Pro release to App, web, and API simultaneously, adding an "expert mode" and native support for the OpenAI Responses API, with targeted adaptation for Codex. Chinese media benchmarks put it within 0.1 points of Anthropic's Claude Fable 5 at roughly 1/60th the cost.
  • Source: ITHome · 2026-08-13
  • Editor's Take: "Close enough, 60x cheaper" is now DeepSeek's entire brand thesis, and it keeps getting harder to argue with.

2. DeepSeek introduces peak/off-peak API pricing, effective August 17

  • Summary: DeepSeek will charge double during peak hours (Beijing time 9:00-12:00, 14:00-18:00) from August 17, with off-peak windows priced at half the peak rate. The move signals a shift from subsidized pricing toward commercial discipline as usage scales.
  • Source: ITHome · 2026-08-13
  • Editor's Take: The "free lunch" era was never going to last. Peak-load pricing is how you manage a model that suddenly everyone wants to run.

3. DeepSeek open-sources V4-Pro-0813 and launches Harness — its answer to Claude Code

  • Summary: Alongside the open-sourced V4-Pro-0813 weights, DeepSeek released a public beta of Harness, its coding-agent framework, with an npm plugin ecosystem. The GitHub repos remain private for now, but the npm packages are live.
  • Source: ITHome · 2026-08-13
  • Editor's Take: DeepSeek isn't just selling tokens anymore, it's building the coding workflow around them. That's the real challenge to Western incumbents.

4. Mistral picks Zhipu GLM-5.2 as its first third-party model: China's models go to Europe

  • Summary: Zhipu's GLM-5.2 becomes the first third-party model on Mistral AI's platform, announced August 11 as part of Mistral's European sovereignty push. Third-party models will run on the same infrastructure, regional controls, and service commitments as Mistral's own.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Chinese open-weight models just got a distribution channel into Europe's enterprise market. Strategic, not just symbolic.

5. WeChat unveils in-house WeLM family: 80B now serving, 617B in development

  • Summary: Tencent WeChat's team revealed WeLM, its efficiency-focused LLM family. The WeLM-80B (80B total, 3B active parameters) is already deployed in WeChat's AI assistant Xiaowei, with a 617B variant in development.
  • Source: ITHome · 2026-08-13
  • Editor's Take: China's most-used app is quietly running its own model at scale. MoE efficiency is the whole point when you serve a billion users.

6. Alibaba open-sources Qwen3.8-2.4T-A95B: first Qwen-Max-grade open weights

  • Summary: Alibaba's Qwen team released weights for Qwen3.8-2.4T-A95B — 2.4 trillion total parameters, 95B active per token, with native 256K context. It's the first time a Qwen-Max-class model has been open-sourced.
  • Source: ITHome · 2026-08-13
  • Editor's Take: The frontier of open-weight models now includes a 2.4T-parameter Chinese model. "Top models stay closed" just took another hit.

7. Cambricon posts 122.6% profit growth, adapts five domestic LLMs, sixth-gen chip in the works

  • Summary: Cambricon reported H1 2026 revenue of ¥5.996 billion (+108.13%) and net profit of ¥2.311 billion (+122.61%). Chairman Chen Tianshi said the company has completed adaptation for GLM, DeepSeek and other top domestic models, with sixth-generation chip architecture under development.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Domestic AI chips are past the "policy story" stage — the numbers are now doing the talking.

II. AI Software & Applications

1. Ant Alipay's AI assistant "Abao" opens beta for Huawei HarmonyOS users

  • Summary: Alipay's AI answer, Abao, is now in open testing for Huawei HarmonyOS users via Huawei's app store (version 12.12.16+). Users summon it by searching "Abao" in the search bar.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Every Chinese super-app now has an AI copilot bolted on. The question is which one becomes the default entry point.

2. iFlytek releases enterprise AI suite covering seven core business scenarios

  • Summary: iFlytek unveiled its enterprise agent matrix on August 12, spanning intelligent management, process IT, supply chain, marketing & sales, R&D & manufacturing, decision analytics, and going-global operations.
  • Source: QbitAI · 2026-08-13
  • Editor's Take: Enterprise agents are where Chinese AI vendors are actually monetizing. The seven-scenario playbook is a land-grab more than an innovation.

3. Study of 6,000+ complaints finds AI coding agents deleting files, generating malicious code (Chinese Team Research)

  • Summary: Researchers at York and Calgary universities analyzed 3,801 Reddit threads on AI coding mishaps (Feb 2023–Mar 2026), finding agents with excessive permissions overwriting files, deleting data, and even generating malicious code. Cursor logged the most incidents.
  • Source: ITHome · 2026-08-13
  • Editor's Take: The lesson is the same as it ever was: the agent is fine, the permissions are the danger.

4. Huawei's MatePad Edge receives HarmonyOS 7 beta with AI voice-clone detection

  • Summary: Huawei began pushing the HarmonyOS 7 beta (7.0.0.102 SP5) to MatePad Edge tablets, adding immersive lighting effects, anti-scam shields, and AI voice-clone detection to guard against voice deepfakes.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Voice-clone scams are surging in China, and putting detection on-device is the right countermeasure. Every platform needs this.

III. Humanoid Robots

1. World Humanoid Robot Games: 2,056 robots from 16 countries, entries up 138%

  • Summary: The 2nd World Humanoid Robot Games opens August 22 at Beijing's "Ice Ribbon" National Speed Skating Oval: 666 teams and 2,056 robots from 16 countries across six continents, a 138% increase in team entries year-on-year.
  • Source: ITHome / CCTV · 2026-08-13
  • Editor's Take: The Olympics of robots is growing 2.4x. China is turning humanoid robotics into a spectator sport — and a supply chain.

2. Xiaomi adds dual mechanical arms and 4cm obstacle clearance to Mijia robot vacuum

  • Summary: Xiaomi's Mijia Robot Vacuum 5 (launched Dec 2025, ¥2,999 list) features dual mechanical arms and 4cm obstacle clearance, and hit a JD.com promo price of ¥1,510 during the site's 815 anniversary sale.
  • Source: ITHome · 2026-08-13
  • Editor's Take: House robots with arms are heading mainstream, and the price war has already started.

IV. Autonomous Driving

1. Pony.ai begins mass production of 4th-gen driverless trucks, targets 1,000 units in three years

  • Summary: Pony.ai has started mass production of its fourth-generation driverless heavy trucks, aiming for 1,000 trucks in operation within three years. The trucking arm began in H2 2018 with a single retrofitted Dongfeng truck.
  • Source: Leiphone · 2026-08-13
  • Editor's Take: Robotaxis grab headlines, but driverless freight is where the unit economics start making sense.

2. Huawei's ADS 5 autonomous driving update pushes to Zhijie S7/R7 via OTA

  • Summary: Huawei's Qiankun ADS 5 update cleared filing and began rolling out to Zhijie "dual-7" models (S7 and R7) on the afternoon of August 13.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Huawei's ADS stack keeps shipping over-the-air. Every OTA is a step toward its software-defines-driving pitch.

3. Stelato G9 and HarmonyOS smart-driving launch event set for August 20

  • Summary: Stelato announced its G9 (pre-sale from ¥439,800) and a broader HarmonyOS smart-driving launch event for August 20 at 14:30. The G9 measures 5377/2050/1897mm with a 3160mm wheelbase.
  • Source: ITHome · 2026-08-13
  • Editor's Take: Huawei's car ecosystem is accelerating release cadence — the G9 is aimed squarely at the premium executive segment.

V. World Models / Physical AI

1. Dyna Robotics unveils DYNA-2: 90% task success trained on 1M+ hours of human video

  • Summary: Embodied intelligence company Dyna Robotics released robot foundation model DYNA-2, trained on over 1 million hours of first-person human video, claiming up to 90% task success. It's positioned as a "World Action Model."
  • Source: ITHome · 2026-08-13
  • Editor's Take: A million hours of "how humans do things" video is the new pretraining data. Robots learning from human motion, not sim reality, is the right bet.

2. Embodied-data startup SCALEFORCE raises two rounds in 40 days for physical AI infrastructure

  • Summary: Yuandian Tech / SCALEFORCE, an embodied-intelligence data infrastructure startup, closed two funding rounds within 40 days (tens of millions of RMB combined), backed by domestic embodied intelligence industry funds and Hengxu Capital.
  • Source: QbitAI · 2026-08-13
  • Editor's Take: Data infrastructure for physical AI is the new picks-and-shovels play. Two rounds in 40 days says investors agree.

EastVoice · Frontier of the Physical World | Daily China AI Briefing for Global Readers

Policy

2 replies

?
Ctrl + Enter to reply
Long Ji
Long JiAug 14

The logic behind trading a 0.1-point performance gap for a 60x price difference is that domestic models are pursuing a differentiation strategy, first penetrating the application layer through cost-performance ratio. However, inference latency is indeed critical, especially in real-time rendering scenarios. Model distillation and quantization still need further optimization.

Yu Yewei
Yu YeweiAug 14

The price gap is indeed exaggerated. My art colleagues previously mentioned that the cost of AI-generated assets was still too high. If it could drop to this level, mass-generating textures for open-world projects would truly be feasible. However, we still need to watch out for shader performance bottlenecks; model inference latency must not impact real-time rendering.