PHYSIX FRONTIER EASTVOICE
2026.09.21 · Monday
Issue 043
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
600B
StepFun's Step 5 Preview sparse MoE total parameters
$290 million
Anew Labs first institutional round funding amount
¥19,999
Agibot Qiyuan Q1 humanoid robot starting price
107
Weekly China venture financing events recorded
Top Stories 3 · dual commentary ｜ Section Picks 23 · one-line takes
StepFun unveils Step 5 Preview; Alibaba releases Qwen-Image-2.1; Unitree sheds ¥250B market cap.
Zhipu AI said its model-serving platform is rolling out a no-content-retention mode for enterprise and developer customers. Once activated, inputs and outputs are not persisted server-side; any user can apply for it.
What it means for you｜Enterprise and developer customers can activate a no-content-retention mode where inputs and outputs are not persisted server-side.
Editor's take｜Privacy guarantees are becoming a billing feature in China's model wars. When pricing converges, the compliance sheet is the new leaderboard.
Alibaba's Qwen team released Qwen-Image-2.1, merging text-to-image and image editing in one model, supporting transparent-image output and up to 10 reference images per generation. The team pitches it as the best quality-per-cost model in its open image series.
What it means for you｜Users get transparent-image output and support for up to ten reference images per generation in one unified model.
Editor's take｜Reference-image conditioning at this scale is how Chinese labs close the gap with closed rivals, one weight dump at a time. Editing and generation in one backbone is what Western labs still ship as two products.
StepFun announced flagship Step 5 Preview, a sparse MoE with 600B total parameters aimed at coding, software engineering, professional work and finance. It already ranks top-three among open models on Artificial Analysis, and StepFun committed to releasing full weights on October 15.
What it means for you｜Developers wait for full weights release on October fifteen after preview ranks top-three among open models.
Editor's take｜A preview that beats released open models is a marketing move, but the October 15 date is the real commitment. Watch whether they actually ship on that day — Chinese labs have been burned on open-source promises before.
23 AI stories worth your time, each with a one-line take.
China Telecom demoed an enterprise AI office suite built entirely on domestic models that runs on one consumer RTX 3090, letting companies keep documents and prompts inside their own network. It enters a summer arms race where big Chinese firms all chase the AI-office entry point.
What it means for you｜Companies keep documents and prompts inside their own network using domestic AI office software on consumer hardware.
One-line take｜The pitch isn't intelligence, it's air gap. Sovereignty sells in China the way low latency sells elsewhere, and a telco bundling models with private lines is structurally hard to undercut.
Shanghai-based APUS said its AI lab published among the earliest independent open-source reproductions of the newly released Jev model, demonstrating sub-second decision latency on domestic models.
What it means for you｜Domestic models demonstrate sub-second decision latency through independent open-source reproductions of the newly released Jev model.
One-line take｜Reproducing a hot model within days is the open-source ecosystem's real speed test. China's labs are no longer just consumers of frontier releases; the turnaround clock now runs in their favor.
At the World Manufacturing Convention in Hefei, MIIT vice minister Xin Guobin said the ministry will push open-source foundation models and vertical-domain models as core tools, with "AI + manufacturing" as the main attack direction for the next phase of industrial policy.
What it means for you｜State-backed enterprises may start writing requirements to run on open weights into tenders following ministry policy statements.
One-line take｜When a ministry names open source in a keynote, procurement follows. Expect state-backed enterprises to start writing "must run on open weights" into tenders.
Frost & Sullivan's 2026 desktop AI agent market report, released September 20, ranks Tencent's WorkBuddy first in China on both the personal and enterprise desktop-agent charts.
What it means for you｜Tencent WorkBuddy ranks first in personal and enterprise desktop-agent charts by leveraging WeChat, Docs, and meeting data distribution.
One-line take｜WeChat, Docs and meeting data all feed one assistant — distribution beats benchmark again. The number one spot in agents in China is really a distribution scoreboard.
Yuanbao app V2.85.0 adds record-while-capturing, internal device audio, custom editing, and one-tap sharing into WorkBuddy and Tencent Docs.
What it means for you｜Meeting transcripts auto-filed into Tencent Docs feed the same corpus that WorkBuddy retrieves from via new recording features.
One-line take｜The recorder is a trojan horse. Every meeting transcript auto-filed into Tencent Docs feeds the same corpus WorkBuddy retrieves from — that's the strategy in one feature note.
Qwen upgraded its minor-protection mode: teen accounts keep photo homework help, AI writing, translation and deep research, while responses are age-filtered, chatbot boundaries are spelled out, and usage-time reminders are enforced.
What it means for you｜Teen accounts access photo homework help, AI writing, translation, and deep research with enforced usage-time reminders and age filtering.
One-line take｜Regulator-facing, almost certainly. But a whitelisted study-only mode is also a wedge: the kid with the Qwen homework habit is the adult with the Qwen subscription.
The domestic model-serving cloud completed its second B+ tranche and a Series C, bringing equity funding this year to roughly ¥2.9 billion / ~$400 million.
What it means for you｜Customers needing someone else to run open-weight models utilize SiliconFlow inference serving cloud services funded by recent rounds.
One-line take｜Inference serving is where China's open-weight glut monetizes. More open models means more customers who need someone else to run them — SiliconFlow sells the shovel after the gold rush gave away the gold.
Anew Labs, spun out of ByteDance's AI drug line in June, closed a $290 million first institutional round at roughly $1.5 billion post — the largest single AI-pharma round in China this year. Its fastest pipeline asset is still preclinical.
What it means for you｜Investors value compute and data engines over molecules as Anew Labs raises $290 million while preclinical assets remain early.
One-line take｜A preclinical company at unicorn pricing tells you investors are valuing the compute and the data engine, not the molecule. That's either visionary or very early Dot-com behavior, and history says you can't tell yet.
Weekly China venture tallies recorded 107 financing events, with a clear tilt toward humanoid-robot parts makers — actuators, dexterous hands, sensors — landing orders inside global OEM supply chains.
What it means for you｜Humanoid-robot parts makers land orders inside global OEM supply chains as component suppliers win over glamorous demo companies.
One-line take｜The boring layer keeps winning. Nobody shorted a component supplier — the boring parts trade beats the glamorous humanoid demo on cash flow.
Unitree Robotics hit ¥444.9B market cap on its first STAR Market day, August 19, then lost ¥250 billion within a month. Embodied-AI startups meanwhile raised about 322 rounds totaling ¥93.5B / ~$13B in H1 2026, a fivefold year-over-year increase.
What it means for you｜Public markets rate commercialization while private markets rate narrative, creating a valuation gap between secondary and primary sectors.
One-line take｜This is the cleanest dataset China's AI boom has produced: public markets rate commercialization, private markets rate narrative, and the gap between the two is the industry's real valuation.
Over six months of record raises — Unitree's listing, Agibot's HK IPO push, Qianxun's single ¥1B round, Xinghaitu's ¥2B B+ — a parallel outflow has gone underreported: founding technologists and commercial leads leaving their own companies, per a GlowStream survey of the sector.
What it means for you｜Founding technologists and commercial leads leave companies during record raises, signaling bearish sentiment despite available funding rounds.
One-line take｜Money arrives, founders exit. The most honest bearish signal in a bubble is who declines to spend the new round at your company.
Dozens of locally backed embodied-AI "data collection centers" proliferated across China; founder criticism from Mech-Mind's Shao Tianlan of deal-stacking has drawn attention, and reporting notes many centers now struggle to prove their training data gets used.
What it means for you｜Data collection centers struggle to prove training data gets used, revealing fiction in dataset economics despite real robots.
One-line take｜Subsidized teleoperation farms producing data nobody trains on is a very specific failure mode of industrial policy. The robots were real; the dataset economics were fiction.
At a closed Hard AI 100 session covered by Huxiu, Chinese AI-application founders described shrinking feature moats — a capability built this quarter gets commoditized by the next model release — and unit economics where usage growth accelerates losses.
What it means for you｜AI-application founders face shrinking feature moats and unit economics where usage growth accelerates losses due to open-source model supply.
One-line take｜This is the applied layer's original sin everywhere, and Chinese founders are saying it out loud because their model supply is open-source and free, which strips away the last excuse: there is no moat but distribution and vertical data.
At its global product launch on September 20, Qiyuan Robotics (backed by "Bumblebee He") priced the Q1 humanoid at ¥19,999, with an Explorer edition at ¥26,999 — consumer-electronics pricing for a full humanoid.
What it means for you｜Consumers can buy a full humanoid robot at sub-$3,000 pricing, turning it into a developer console rather than museum piece.
One-line take｜Sub-$3,000 is the price point where a humanoid becomes a developer console instead of a museum piece. Whatever Q1 does at launch, the hackable install base is the asset.
Ifanr's hands-on with the ¥19,999 Qiyuan T1 argues the company is doing the opposite of the industry: instead of promising future domestic robots, it is shipping companion, telepresence and camera-tracking functions people will actually pay for today.
What it means for you｜Buyers pay for companion, telepresence, and camera-tracking functions today instead of waiting for promised future domestic robot capabilities.
One-line take｜Every embodied-AI deck promises the kitchen; Qiyuan is selling the living room. Beachhead companionship into later utility is at least a plan, which is more than most demo-videos companies have.
Faraday Future held a September 19 launch event in the US unveiling its "quad-core full-intelligence" EAI robot line, the All-New Futurist, Master Mini and FX Aegis series — nine configurations, the most expensive priced above ¥920,000.
What it means for you｜Customers choose from nine EAI robot configurations priced above ¥920,000 from an EV company lacking scale shipping track record.
One-line take｜An EV company that never shipped cars at scale is now shipping the priciest humanoid SKU on the board. The market should price the track record, not the configurator.
Dongfeng's chief intelligent-technology engineer Zhang Zhenlin said humanoid robots will begin sorting and quality-inspection work at Dongfeng factories this October, with small-batch trials by year-end and a stated goal of human-parity capability by end of 2027. The company's robot-dog program is also in development.
What it means for you｜Factory workers see humanoids begin sorting and quality-inspection tasks this October with small-batch trials by year-end.
One-line take｜A state automaker deploying its own robots, on its own timeline, for its own QA line is the least glamorous and most credible humanoid story in China. Self-hosting beats selling.
GeekPark profiled Levink's self-evolution approach, in which robots improve manipulation from their own deployed operation data rather than centrally curated teleoperation datasets.
What it means for you｜Robots improve manipulation from deployed operation data rather than centrally curated teleoperation datasets, challenging data center business models.
One-line take｜Self-evolving fleets vs. data farms is the embodied-AI training-strategy fork of the year. Whichever compounds on deployment data wins the cost curve — the data centers' business model only works if the robots don't.
JD.com launched "Qixian Dachu", a family healthy-meal service built around a cooking robot, opening pre-orders in Beijing, Shanghai, Guangzhou and Shenzhen with a ¥19,888 founding-member price.
What it means for you｜Home users subscribe to recurring meal-kit services around cooking robots, with JD buying kitchen entry via hardware charges.
One-line take｜The robot is the marquee; the recurring meal-kit subscription is the business. JD is buying its way into the kitchen through the back door of a hardware one-time charge.
Harmony Intelligent Mobility Alliance opened pre-orders for a new battery variant of the Stelato RX crossover: lidar-equipped cabin trim from ¥269,800, and a trim with L3 autonomous-driving architecture from ¥329,800.
What it means for you｜Car buyers treat L3 autonomous-driving architecture trim level as regulatory readiness when selecting Stelato RX variants.
One-line take｜Huawei's pricing sheet now lists an L3 badge as a trim level. Whether that's regulatory readiness or marketing, buyers will treat it as the former.
Beijing Auto Works unveiled studio photos of the Xingtan 5X off-road-styled SUV and officially confirmed it carries Huawei's Qiankun driver-assistance stack.
What it means for you｜Drivers rely on Huawei Qiankun stack in BAIC Xingtan 5X SUVs as licensing expands across vehicle brands.
One-line take｜Huawei's answer to Android is licensing the driving stack to everyone who can't build one. The more badges, the stronger the moat — as long as the cloud keeps paying.
A widely shared Guokr piece catalogs owners' frustration with ADAS that slams the brakes on open roads with no visible cause — "the hardest part of smart driving was never making the car go, it's teaching it to stop" — as phantom braking complaints cluster across Chinese forums.
What it means for you｜Owners experience frustration with ADAS slamming brakes on open roads without visible cause, impacting trust in smart driving products.
One-line take｜Conservative braking is statistically safe and experientially enraging, and trust is the actual product being sold here. Recall rates won't capture churn of this.
On the final WorldArena 2.0 leaderboard, Coscene's Boundless world model took second globally on Track 1 video quality while ranking first on background consistency and JEPA similarity — its claim to long-horizon scene coherence leadership.
What it means for you｜World model users benefit from long-horizon scene coherence leadership demonstrated through background consistency and JEPA similarity metrics.
One-line take｜Consistency metrics are where world models get honest — visual fidelity flatters, coherence over minutes betrays. A Chengdu company leading those is worth more attention than the demo reels suggest.
📖 The full EastVoice issue on the Chinese forum
📈 Physix Frontier · Alpha — the investor's read
Headline teardowns · curated signals · market moves. Five minutes to find the alpha of the AI era →