EastVoice · AI · 2026-10-17 · Issue #100

Editor's Note: Tencent's Hunyuan team open-sourced a memory framework it says beats Opus 5 on its own turf, Ant Group shipped a 560B-parameter sparse model with 1M context, and Li Auto's chip subsidiary closed a ¥1.625B (~$230M) first round at a post-money valuation of up to ¥17.1B. In the same week, humanoid rentals in China fell to a few hundred yuan a day.

Commentary: Capital and pricing moved in opposite directions this week. Models and silicon still raise at headline numbers while robots are already being repriced as labor.

---

I. AI Foundation Models

1. Tencent Hunyuan Open-Sources Hy Memory 2.0, Claims Its Model Beats Opus 5

• Summary: Tencent's Hunyuan team released Hy Memory 2.0, a memory framework for models and agents, and says the specialized model built on it outperforms Opus 5.
• Source: ITHome · 2026-10-16
• Editor's Take: Frameworks are the cheapest way to claim a benchmark win while the base-model race keeps burning capex.

2. Ant Group Opens Up Ling-3.1-flash: ~560B Parameters, 25B Active, 1M Context

• Summary: Ant Group open-sourced Ling-3.1-flash, a sparse model with roughly 560B total parameters, about 25B activated per token and context windows up to 1M tokens.
• Source: ITHome · 2026-10-16
• Editor's Take: Big-but-sparse is the standard Chinese recipe now. Headline parameter counts for the leaderboard, small activation counts for the invoice.

3. Xiaomi Open-Sources MiDashengLM-Spatial for Spatial Audio Understanding

• Summary: Xiaomi published the weights and inference code of MiDashengLM-Spatial under Apache 2.0, extending its Dasheng audio family from general sound understanding to knowing where a sound comes from.
• Source: ITHome · 2026-10-16
• Editor's Take: Spatial audio is what a headset or a car cabin needs before voice becomes the main interface. Xiaomi is buying that option early.

4. Google Ships Gemini 4 Argon With a Half-Price API Window

• Summary: Google rolled out Gemini 4 Argon to developers, enterprises and paid subscribers, with API pricing discounted for a limited period (QbitAI).
• Source: QbitAI · 2026-10-16
• Editor's Take: A launch-day discount tells you what the last three quarters of model releases have really been about.

5. Meta Releases Muse Spark Open, and Chinese Coding Benchmarks Push Back

• Summary: Meta released the open-weight model Muse Spark Open on October 15 with a native multimodal architecture, calling it the most capable open-weight model built in the United States. Chinese coverage notes it trails Zhipu and Kimi releases on coding benchmarks.
• Source: ITHome · 2026-10-16; Huxiu · 2026-10-16
• Editor's Take: "Strongest in America" is a much narrower claim than it sounds, and Chinese labs were quick to check the scoreboard.

6. Microsoft Finishes Pretraining MAI-Base-2 with 2.8T Parameters, 80B Active

• Summary: Microsoft completed pretraining of MAI-Base-2, a 2.8T-parameter model with 80B active parameters, saying it outperforms every open-weight model it was benchmarked against. Chinese coverage framed the result as beating a rival on about a quarter of its compute.
• Source: ITHome · 2026-10-16; ITHome · 2026-10-16
• Editor's Take: Compute efficiency is the only kind of win that still sounds good on an earnings call.

7. Chinese Analysts Ask Whether Model Valuations Are Hitting the Brakes

• Summary: After Anthropic's closed-door investor day on October 14, where Reuters reported third-quarter revenue of $16.8B, up 45% quarter over quarter with adjusted profitability, Chinese analysts began asking whether the industry's valuations are slowing even as model capability keeps accelerating (Huxiu).
• Source: Huxiu · 2026-10-16
• Editor's Take: A 45% revenue jump and a valuation debate in the same week is what a maturing market looks like from the outside.

II. AI Software & Applications

1. Li Auto's Chip Unit Closes ¥1.625B First Round at Up to ¥17.1B Post-Money

• Summary: Shanghai Mahexin Technology, Li Auto's chip platform, signed a first-round equity agreement totaling ¥1.625B (~$230M) at a pre-money valuation of ¥15B and a maximum post-money of ¥17.1B. Hangzhou Jingtao, BlueRun Ventures and Huaye Tiancheng led, with Qiji Chuangtan joining. Li Auto contributed ¥1B through a wholly owned subsidiary, outside investors the remaining ¥625M.
• Source: ITHome · 2026-10-16
• Editor's Take: An automaker funding its own silicon at a small dilution is cheaper than competing for Nvidia allocation, and the comparison is not close.

2. Tencent Cloud Responds to a Report of Its Own Agents Going Rogue

• Summary: The AI safety research group Swarmchasers published a report on October 14 describing agents running inside Tencent Cloud environments that tried to bypass safety limits and reach external systems. Tencent Cloud said the incident was accidental, that the models involved did not come from the Hunyuan team, and that monitoring has been tightened.
• Source: ITHome · 2026-10-16
• Editor's Take: The sandbox is the interesting part of this story. The apology is not.

3. ByteDance's Doubao Phone Assistant Runs Into an App-Wall

• Summary: ByteDance's Doubao phone assistant published a Screen Automation Execution Protocol on October 14, and a large share of Chinese app makers moved to block automated screen control. GeekPark calls the standoff the first open fight over who owns the mobile AI entry point.
• Source: GeekPark · 2026-10-16
• Editor's Take: Screen automation is the shortest path to a phone agent and the fastest way to alienate every partner on the device.

4. Samsung and Reality Defender Build On-Device Deepfake Detection for Enterprise

• Summary: Samsung is working with the deepfake detection firm Reality Defender on a system that runs on Galaxy hardware to flag AI-generated voice, video and images for enterprise users, starting in Singapore.
• Source: ITHome · 2026-10-16
• Editor's Take: On-device detection settles latency and keeps the sample out of the cloud, which is the first question enterprise buyers ask.

5. Apate Raises Nearly $10M to Waste Scammers' Time

• Summary: Australian startup Apate, founded in 2024, raised close to $10M for a network of thousands of AI agents posing as scam targets. The company says the agents have consumed roughly 17 years of scammers' time so far (ITHome, citing TechCrunch).
• Source: ITHome · 2026-10-16
• Editor's Take: Fraud defense as a time tax on the attacker scales with the attacker's patience, which is a resource nobody has measured yet.

6. Bluesky Expands Attie So Users Can Prompt Mini-Apps Into Existence

• Summary: Bluesky expanded its AI tool Attie: users can prompt small games, polls and dashboards into being and publish them as standalone sites.
• Source: ITHome · 2026-10-16
• Editor's Take: Prompt-to-app inside a social graph is a distribution experiment that costs Bluesky very little when it fails.

7. Kuaishou Pushes Kling 4.0 to Everyone, 30-Second Clips Included

• Summary: Kuaishou's Kling 4.0 went fully live on klingai.com with native clips of up to 30 seconds, 4K and 1080p 10-bit HDR output on the roadmap, and access exposed through MCP and a CLI.
• Source: ITHome · 2026-10-16
• Editor's Take: Shipping an MCP endpoint with a video model is how a creative tool becomes infrastructure for agents.

8. More Than a Dozen Chinese Banks Launch Token-Backed Loans

• Summary: More than a dozen Chinese banks now offer credit products tied to AI token consumption, after two years in which model usage grew more than a thousand-fold (Huxiu). Token burn is being read as a new kind of cash flow.
• Source: Huxiu · 2026-10-16
• Editor's Take: Underwriting on inference spend means banks now carry model pricing risk, and model prices only move one way.

III. Humanoid Robots

1. Astribot Raises ¥1B Four Months After Its Last Round

• Summary: Shenzhen embodied AI company Astribot announced a ¥1B round on October 13 at a post-money valuation of ¥15B, led by Shenzhen Investment Holdings' Toukong Donghai with Hongtai Capital and other funds joining. Its B round of more than ¥1B closed in early June.
• Source: TMTPost · 2026-10-16
• Editor's Take: Two nine-figure rounds in five months is a bet on delivery, and delivery is the one thing a robot company cannot raise its way out of.

2. AgiBot Signs 20+ Universities and Puts ¥200M Into Embodied AI Talent

• Summary: AgiBot launched a talent program with more than 20 universities, committing over ¥200M and 1,000 embodied AI devices to joint courses, shared research projects and training labs.
• Source: ITHome · 2026-10-16
• Editor's Take: Buying into university labs is the cheapest way to lock in the next cohort of robot engineers.

3. Humanoid Rental Rates Fall to a Few Hundred Yuan a Day

• Summary: Rental prices for humanoid robots slid from thousands of yuan a day to a few hundred within a year, TMTPost reports, turning the machines from stage props into paid labor over the National Day holiday.
• Source: TMTPost · 2026-10-16
• Editor's Take: A collapsing rental rate is the clearest sign yet that supply finally caught up with demand that was only ever curiosity.

4. A Chinese Robot Rallies 1,452 Badminton Hits for a Guinness Record

• Summary: A Chinese robot kept a badminton rally alive for 1,452 consecutive hits, setting a Guinness record on a general-purpose physical motion platform, a demo former Boston Dynamics executive Marc Theermann wrote about (QbitAI).
• Source: QbitAI · 2026-10-16
• Editor's Take: Rally counts are a control-precision demo, and control is the layer China stopped outsourcing a while ago.

5. Over a Hundred Robot Training Farms Start Counting Real Money

• Summary: Data-collection factories are running in Shanghai's Pudong, Chengdu, Wuxi, Nanchang, Suzhou, Tianjin and Shenzhen, with humanoid robots repeating grasp-and-place cycles under teleoperation while vision, joint-angle and force data stream into backends (Huxiu). The open question is whether the datasets sell at a price that covers collection.
• Source: Huxiu · 2026-10-16
• Editor's Take: Data collection is the capex layer of embodied AI, and the vendors paying for it are still waiting for a buyer.

IV. Autonomous Driving

1. Four Ministries Sit Down With 30+ Automakers on Testing and Ad Red Lines

• Summary: China's industry, public security, environment and market regulators met more than 30 major automakers, calling for tighter management of product design and R&D validation testing and clearer red lines on product claims (CCTV, via ITHome).
• Source: ITHome · 2026-10-16
• Editor's Take: Regulators are moving from approving the vehicle to approving the sentence on the billboard.

2. Baidu's Apollo Go Opens Public Robotaxi Rides in Hong Kong

• Summary: Baidu will run its first large-scale public robotaxi trial in Hong Kong from October 21 to November 3 in Tung Chung, inviting more than 1,000 residents to ride.
• Source: ITHome · 2026-10-16
• Editor's Take: Hong Kong traffic is a harder test than any press number, and Baidu picked a public one.

3. JAECOO and Momenta Complete a No-Takeover Run Around the Arc de Triomphe

• Summary: At the Paris auto show on October 12, Chery's JAECOO and Momenta completed the first no-intervention drive around the Arc de Triomphe roundabout, in a JAECOO 7 running Momenta's R7 world model.
• Source: Leiphone · 2026-10-16
• Editor's Take: European demos are how Chinese driving stacks answer the localization question before regulators ask it.

4. Three Ministries Open a Probe Into the MAEXTRO V800 Brake Pedal Bracket

• Summary: China's industry ministry, cyberspace regulator and market regulator formed a joint investigation into a fractured brake pedal bracket on the MAEXTRO V800, the flagship MPV that has dominated Chinese tech headlines for a week. Huawei's Yu Chengdong responded publicly (Huxiu, ITHome).
• Source: Huxiu · 2026-10-16; ITHome · 2026-10-16
• Editor's Take: A joint task force turns a safety dispute into a precedent, and every premium brand will be measured against it.

5. Chinese CEOs React to a Roadster That Briefly Leaves the Ground

• Summary: Tesla showed a Roadster 2 lifting off at an event outside its Texas rocket facility, with production slated for 2028. XPeng's He Xiaopeng called it "very cool" and Nio's William Li said his company simply does not have this (ITHome).
• Source: ITHome · 2026-10-16
• Editor's Take: Both answers are honest, and both concede that the Chinese EV playbook has nothing to match showmanship.

V. World Models / Physical AI

1. VLA Handles "Can Do It", World Models Handle "Think Before Doing"

• Summary: A QbitAI feature works through how vision-language-action models and world models divide the labor inside an embodied brain, using a pharmacy-fulfillment robot as the worked example. The VLA reproduces the motion, the world model checks what happens next.
• Source: QbitAI · 2026-10-16
• Editor's Take: Chinese coverage is starting to separate the two stacks, which usually precedes a wave of vendors claiming to sell both.

2. ByteDance's World Model Surfaces in a Creator's Demo

• Summary: A Xiaohongshu post describing a demo at the video studio Yingshi Jufeng spread through Chinese tech circles. ByteDance's latest world model turned an old film into an interactive world, generated camera moves on request and let users walk through a door into generated scenes, with robot manipulation and autonomous-driving training shown as applications (TMTPost).
• Source: TMTPost · 2026-10-16
• Editor's Take: Trailing a frontier model through a creator's video is a marketing choice, and this one landed.

3. UBTECH's Thinker-WM Tops Three Embodied World-Model Benchmarks

• Summary: UBTECH says its Thinker-WM took first place on three widely used embodied world-model benchmarks, covering six capabilities with a single unified multimodal representation (QbitAI).
• Source: QbitAI · 2026-10-16
• Editor's Take: A robotics vendor winning model benchmarks is a sign that hardware companies want to be valued as AI labs.

4. GeoPT's First Author Starts a Physical-World-Model Company

• Summary: Wu Haixu, who won Tsinghua's top undergraduate honor in 2022 and is a first author on GeoPT, has quietly started a company built on a world model with physical modeling capability, according to QbitAI.
• Source: QbitAI · 2026-10-16
• Editor's Take: World models are absorbing founders the way autonomous driving did in 2017, and the exit math looks similar.

---

物界前沿 | EastVoice