Physix Frontier badgePhysix Frontier
PublishChinese | EN
Community Discussion · General

EastVoice · AI · 2026-10-20 · Issue #103

EastVoiceEastVoiceOct 202026/10/19 505 views

Editor's Note: China's model labs spent Monday trading margin for position. MiniMax launched M3.1 Flash at 3x the speed and a 70% lower price, DeepSeek put a ¥5 per 1,000 calls tag on its own search API, and Tencent's Hunyuan team published a self-improving agent stack built on Alibaba's base model. On the same day, MIIT put a number on the robot supply chain.

Commentary: Cheap tokens stopped being a differentiator this year. What is left is distribution.

I. AI Foundation Models

1. MiniMax Ships M3.1 Flash With 3x Speed and 70% Lower Price

  • Summary: MiniMax released M3.1 Flash on Monday, claiming 3x faster inference and a 70% price cut against the previous generation. The model had been serving traffic anonymously as "Space Bunny," processing more than 200 trillion tokens before the launch.
  • Source: ITHome · 2026-10-19
  • Editor's Take: Price cuts at this pace are a land grab. The margin gets paid later, by whoever is still standing.

2. DeepSeek Prices Its Own Search API at ¥5 per 1,000 Calls

  • Summary: DeepSeek published pricing for its Search / Fetch endpoints at ¥5 (~$0.75) per 1,000 successful requests, with failed, empty and timed-out calls billed at zero. Billing starts November 1. The interfaces had been free inside Harness, Claude Code and Codex.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Roughly a tenth of Tavily's price, and the "no result, no charge" clause puts the risk on DeepSeek's side of the table. That is how you buy developers.

3. Tencent Hunyuan Builds a Self-Improving Loop on a Rival's Base Model

  • Summary: Tencent's Hunyuan Frontier team published Terminal-RSI, a recursive self-improvement pipeline that lets a model rewrite its own harness, system prompts, toolchain and skill library. The work sits on Alibaba's base models rather than Tencent's own.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Tencent building openly on a competitor's weights says more about Chinese open-source dynamics than any benchmark chart.

4. Arena's Week 42 Chart Puts Chinese Models Mid-to-Upper Tier

  • Summary: Arena's blind-test ranking for October 12–18 shows a stable top tier. kimi-k3-max entered the coding leaderboard's top ten for the first time, with mimo- and others holding ground across multiple categories.
  • Source: ITHome · 2026-10-19
  • Editor's Take: A top-ten coding slot is the number that matters. Coding is where Chinese labs have been most exposed.

5. Kimi and DeepSeek Have Stopped Looking Alike

  • Summary: TMTPost argues the two most-watched Chinese labs diverged through 2026. Both closed large rounds and wrote long context, coding and agents into their roadmaps, but their product and research priorities now point in different directions.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Convergence was a fundraising story. Divergence is a business story.

6. Moonshot AI Weighs a Hong Kong Listing at a Reported $50 Billion Valuation

  • Summary: Moonshot AI raised a pre-IPO round at roughly $50 billion and is targeting a Hong Kong listing in the first quarter of 2027, according to Bloomberg.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: A private market can buy a long option on AGI. A public market will not.

7. CITIC Securities Turns Bullish on Domestic Foundation Models

  • Summary: CITIC Securities published a note arguing that Harness could become the link that takes DeepSeek from model supplier to platform ecosystem, converting open-source reach and engineering strength into user base and commercial pull.
  • Source: ITHome · 2026-10-19
  • Editor's Take: Sell-side notes on open-source strategy usually arrive after the fact. This one landed early.

8. ByteDance Pushes Back on the "Half of China's Compute Makes Short Dramas" Claim

  • Summary: After a listed company filing suggested 55% of Chinese AI token consumption goes to short dramas and video generation, ByteDance's Doubao lead Liu Xing said only about 10% of ByteDance's total compute goes to video-model training and inference.
  • Source: ITHome · 2026-10-19
  • Editor's Take: The fight is over what counts as useful compute. Everyone is defending their own denominator.

II. AI Software & Applications

1. Qwen Wins Users Refunds, but Platforms Keep Agents Out

  • Summary: A Taobao user recovered ¥52 after Alibaba's Qwen agent found a historical order and filed a price-protection claim on his behalf. In the same week, more than 150 mainstream apps closed their doors to screen-automation agents under the protocol Doubao published.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: An agent can already win you a refund. It still cannot get past a login screen.

2. Doubao Becomes ByteDance's Trunk, and Four Agent Lines Fall In

  • Summary: TMTPost traces the sequence: Feishu's product team merged into Doubao in July, hotel orders routed through the Doubao entry point began carrying a ~12% combined service and payment fee in August, and four agent lines have since reported up into the same structure.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: The entry point was never the prize. The transaction fee was.

3. Xiaohongshu Merges Its AI Unit Into Core Infrastructure

  • Summary: Xiaohongshu told staff on October 16 that its first-level AI department Dots will be merged with the core infrastructure team, and folded its two in-app answer products under a single Dian Dian brand.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Content platforms are learning that an AI answer layer belongs closer to infrastructure than to editorial.

4. AliExpress Starts Making Angel Investments in Hardware Brands

  • Summary: AliExpress opened a program that backs AI hardware startups before the brands exist, targeting devices built to sell overseas — from voice recorders and smart glasses to companion robots.
  • Source: Leiphone · 2026-10-19
  • Editor's Take: The marketplace moved upstream. Funding a brand is cheaper than bidding for it later.

5. Biren Raises a Third Time in Ten Months While Its Stock Slides

  • Summary: Biren Technology has now raised three times since listing in Hong Kong on January 2: HK$6.2 billion net at IPO, HK$7.04 billion in a July placement, and about HK$4 billion in a discounted placement on October 8.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Three raises in ten months is not a growth story. It is a cash-burn disclosure with a share price attached.

6. Volcano Engine Funds Premium AI Short Dramas at Up to ¥200,000 a Title

  • Summary: ByteDance's Volcano Engine launched an incentive program for AI-made short dramas built with Seedance and published on the Hongguo app. Qualifying titles ranked monthly in the app's AI drama top 50 can receive up to ¥200,000 (~$28,000) each.
  • Source: ITHome · 2026-10-19
  • Editor's Take: Paying per title bootstraps a content format fast. It also prices the format at zero the day the subsidy stops.

7. Manus's Cue Has an Email, a Phone Number and a Wallet — and Still Isn't Always On

  • Summary: A week of hands-on testing with Cue, launched by Manus on September 28, shows each agent holding its own email address, phone number, digital wallet and dedicated machine, able to keep working after the user logs out. TMTPost found it useful for scheduling and email triage, but not yet dependable around the clock.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: Identity was the easy part. Uptime is the product.

III. Humanoid Robots

1. China Built More Than 40,000 Humanoid Robots in the First Half

  • Summary: At a State Council Information Office briefing, MIIT said humanoid robot output passed 40,000 units in the first half of 2026 and is on track to exceed 100,000 units for the full year.
  • Source: ITHome · 2026-10-19
  • Editor's Take: 40,000 units of output with almost no consumer market to absorb them. The inventory question is coming for someone.

2. Unitree Jumps as a National Standard Lands, and Two Banks Disagree

  • Summary: Unitree rose as much as 18% intraday on Monday, touching ¥508.99 and pushing market capitalisation back above ¥200 billion, days after a humanoid robot national standard was published. Two international investment banks have since published opposite conclusions on the stock.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: A standard is a floor, not a forecast. The disagreement between the two banks is the honest signal.

3. Unitree Robots Reach the Factory Floor and Find It Harder Than the Stage

  • Summary: At the Shanghai industrial fair, two Unitree G1 units worked a roughly 30-square-metre booth — one moving parcels to target positions, the other driving an electric screwdriver into threaded holes. Reporters found the gap between demo and full shift is mostly a repetition problem.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: Grasping a parcel once is a demo. Grasping it 10,000 times is a product.

4. Humanoid IPOs Queue Up While the Technology Runs Late

  • Summary: Huxiu's survey of the humanoid IPO pipeline finds valuations floating ahead of demonstrated capability, with investors pricing companies that have not yet shipped commercial deployments.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: Second-tier robot listings are being priced off Unitree's chart, not off their own revenue.

5. A ¥22 Billion Robot Unicorn Cannot Sell Stock at Half Price

  • Summary: TMTPost reports an investor who bought into an embodied-AI startup at a ¥3 billion post-money valuation in January tried to exit at ¥12 billion — half the company's current ¥22 billion mark — and found no buyer.
  • Source: TMTPost · 2026-10-19
  • Editor's Take: The bid-ask gap on paper valuations is now the sector's clearest risk metric.

IV. Autonomous Driving

1. China Opens Consultation on a National Vehicle Networking Standards Plan

  • Summary: MIIT published a draft 2026 edition of the National Intelligent Connected Vehicle Technology Standards System guide for public comment, running October 20 to November 18, with the goal of a complete system in place by 2030.
  • Source: ITHome · 2026-10-19
  • Editor's Take: Standards are how China schedules its own autonomous driving market. The 2030 date is the timetable investors should be reading.

2. Zhijie Set to Win Its Own Manufacturing Licence

  • Summary: Anhui Zhijie New Energy Vehicle appeared in MIIT's 412th batch of road motor vehicle manufacturer announcements, a step toward independent production qualification. Owners may soon see a Zhijie badge where a Chery badge used to be.
  • Source: Huxiu · 2026-10-19
  • Editor's Take: Cutting the apron string from Chery matters more for how the business is valued than for how the car is built.

3. Alibaba's DAMO Academy Debuts an Automotive-Grade RISC-V Processor

  • Summary: At the RISC-V China Summit in Shenzhen, DAMO Academy unveiled the XuanTie E908A automotive-grade enhanced processor and published its automotive product roadmap for the first time.
  • Source: ITHome · 2026-10-19
  • Editor's Take: RISC-V enters the vehicle through its least glamorous door — the chips cheap enough to be worth replacing.

4. Li Auto Hires a Founding Member of Huawei's ADS Team

  • Summary: Yang Xuguang, who led the prediction module inside Huawei's founding ADS team, has joined Li Auto, according to Leiphone's New Auto Drive report. He worked alongside the team leads for dynamic and static perception.
  • Source: ITHome · 2026-10-19
  • Editor's Take: Talent movement between Chinese ADAS teams is now the fastest way to track roadmap changes.

V. World Models / Physical AI

1. A World-Model Researcher Says the Bottleneck Is No Longer the Architecture

  • Summary: In a World Model 50 interview, robotics researcher Jia Xiaosong argues that with VLA and WAM architectures appearing almost monthly, the binding constraint has shifted away from model design toward data and evaluation.
  • Source: Leiphone · 2026-10-19
  • Editor's Take: Architecture papers are cheap now. The teams that win will be the ones that own the data loop.

2. Nanjing Hosts China's First Physical AI and World Model Conference

  • Summary: The 2026 Physical AI & World Model Conference (PAIWM) opened in Nanjing on October 19 and runs through October 22, with nine parallel tracks spanning healthcare world models, industrial world models, autonomous driving and embodied intelligence.
  • Source: QbitAI · 2026-10-19
  • Editor's Take: Naming a conference after a technical paradigm is a bet that the paradigm outlives the funding cycle.

3. Gewu Tongzhi Raises Nearly ¥100 Million for Physics-Inversion World Models

  • Summary: Hangzhou-based embodied-brain startup Gewu Tongzhi closed a nearly ¥100 million angel round, with backing from state-linked funds, a provincial embodied-AI vehicle and several listed companies. The money is earmarked for large-scale pretraining of physics-inversion world models.
  • Source: ITHome · 2026-10-19
  • Editor's Take: State funds writing angel cheques into world models tells you where industrial policy expects the next stack to sit.

EastVoice | 27 items: 8 foundation models, 7 software, 5 humanoid, 4 autonomous driving, 3 world models / physical AI.

General

2 replies

?
Ctrl + Enter to reply
Shen Tou
Shen TouOct 20

DeepSeek's "no result, no charge" is the real killer move — it's basically trial-and-error for developers before pricing.

Old Ye from BCG
Old Ye from BCGOct 21
Reply to Shen Tou

Agreed, but the key is that line in the body: "failures, empty results, and timeouts are all billed as zero" — it shifts the "cost of trial and error" from the developer back onto themselves, essentially buying call volume with gross margin.