PHYSIX FRONTIER EASTVOICE
2026.10.14 · Wednesday
Issue 060
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
¥100B (~$14B)
DeepSeek's latest raise doubles from planned ¥50B
$50B
Moonshot AI reportedly valued at $50B
¥2.3B
Infinigence AI Series C raises ¥2.3B
3GB
ModelBest MiniCPM-V 4.7 runs in 3GB memory
Top Stories 3 · dual commentary ｜ Section Picks 26 · one-line takes
DeepSeek's Round Doubles to ~¥100B, Moonshot AI Reported at $50B, and ByteDance's Seed Unit Gets Its Own Model Safety and Alignment Team.
DeepSeek's latest raise has grown from a planned ¥50B to about ¥100B (~$14B) in under two weeks, according to people involved in the deal, with the pre-money valuation still around ¥500B (~$70B).
Editor's take｜Nobody needs two weeks to double a round because demand appeared. The price was never the constraint on this deal, and investors know it.
Source: ITHome (2026-10-13)
ByteDance has set up a "Seed Model-Alignment & Preparedness" department, led by Yan Lin and reporting to Seed head Wu Yonghui, with the existing Responsible AI team and part of the Chat team moving into it.
Editor's take｜Safety work has been pulled out of a side project and given its own line item and headcount. Chinese labs are now writing the same org chart as their US peers, which is the clearest sign yet that scale has arrived.
Source: ITHome (2026-10-13)
ModelBest released MiniCPM-V 4.7 in 35B-A3B and 1B sizes, the first MoE entry in the MiniCPM-V line. The company claims 13+ tokens/s decoding on hardware with only 3GB of memory.
What it means for you｜A model that runs inside the RAM a phone already has reaches a market no data center can serve.
Editor's take｜The parameter count is marketing. The memory budget is the product. A model that runs inside the RAM a phone already has reaches a market no data center can serve.
Source: ITHome (2026-10-13)
26 AI stories worth your time, each with a one-line take.
At a media briefing in Shanghai, Huawei's Xu Songyu said the Kirin 9050 series runs a 30B-parameter MoE language model directly on the handset, calling it an industry first and the opening of a ten-year chip roadmap.
What it means for you｜If it survives thermal throttling in a real pocket, the case for cloud-only assistants gets much weaker.
One-line take｜Shipping an MoE on a phone is a memory-bandwidth stunt before it is a model achievement. If it survives thermal throttling in a real pocket, the case for cloud-only assistants gets much weaker.
Source: ITHome (2026-10-13)
Moonshot AI has reportedly closed a final round valuing it at about $50B and plans to list in Hong Kong in Q1 2027, roughly a month after filing confidentially with the HKEX.
What it means for you｜The clock on Chinese foundation labs just got a lot louder.
One-line take｜A $50B mark and a listing two quarters later tells you investors want a return, not a research project. The clock on Chinese foundation labs just got a lot louder.
Source: TMTPost (2026-10-13)
AI infrastructure provider Infinigence AI raised ¥2.3B in a Series C, taking cumulative funding to about ¥4.3B, a record for China's AI-native infrastructure sector on both counts.
What it means for you｜Every model that ships makes this layer more defensible, not less.
One-line take｜Infra keeps winning the funding race while model labs win the headlines. Every model that ships makes this layer more defensible, not less.
Source: QbitAI (2026-10-13)
Liang Wenfeng has been meeting Chinese embodied-AI founders, and DeepSeek has reportedly agreed to lead a seed round of several tens of millions of dollars in Winverse AI, which was founded by former Tencent Hunyuan language lead Hu Han, at a post-money valuation near $400M.
One-line take｜This is DeepSeek's second robotics move after an August stake in Unitree worth about ¥141M. A model lab buying its way into bodies is a bet that the next benchmark is not text.
Source: ITHome (2026-10-13)
A team led by Fudan University's Qiu Xipeng, behind the open MOSS line, has raised over ¥1B two years after founding, backing work on models that can listen, watch and respond to a changing environment.
One-line take｜University labs in China now raise at the scale of later-stage companies. That is part of why local model quality has moved faster than most forecasters expected.
Source: GeekPark (2026-10-13)
Alibaba's Qwen has opened in-app applications for Qwen Go!, a personal agent that can file flight-delay insurance claims, arrange car maintenance, book sports sessions and compare prices. Early users reported a ¥70 delay payout and a ¥52 Taobao price-protection refund handled from start to finish.
What it means for you｜An agent that reliably claws back small amounts of money is worth more than one that writes poetry about it.
One-line take｜The demo that matters is the ¥52 refund, not the model card. An agent that reliably claws back small amounts of money is worth more than one that writes poetry about it.
Source: ITHome (2026-10-13)
Alibaba's supply-chain platform 1688 is opening the Accio AI commerce workbench to more than 100M annual active buyers and over 1M source factories, letting buyers procure and sellers run storefronts through agents.
What it means for you｜If agents work anywhere, it is here, and 1688 owns the inventory.
One-line take｜Wholesale procurement is thin-margin, high-repetition work. If agents work anywhere, it is here, and 1688 owns the inventory.
Source: Leiphone (2026-10-13)
Model maker MiniMax has opened its first Tmall flagship store, listing M Plan subscriptions in Go, Explore and Build tiers with monthly and annual billing.
What it means for you｜Putting token plans on a shopping platform turns model access into a retail SKU.
One-line take｜Putting token plans on a shopping platform turns model access into a retail SKU. That is what a commodity looks like from the inside.
Source: ITHome (2026-10-13)
ByteDance merged TraeWork and TraeCode into a single TRAE product on October 9, repositioning it from an AI coding tool to an AI development platform where agent and IDE modes switch without friction.
What it means for you｜Consolidating this early is a quiet admission about how thin the coding-agent market is.
One-line take｜Two products became one platform because the buyer does not care which mode they are in. Consolidating this early is a quiet admission about how thin the coding-agent market is.
Source: TMTPost (2026-10-13)
Tencent will retire QClaw, one of the early desktop agents, on December 24, less than a year after launch.
What it means for you｜Desktop agents are being absorbed into bigger suites before any of them built a real user base.
One-line take｜QClaw arrived with noise and is leaving almost without it. Desktop agents are being absorbed into bigger suites before any of them built a real user base.
Source: TMTPost (2026-10-13)
Butterfly Effect, the parent of Manus, has closed more than $500M in new funding, reportedly at a target valuation of $4B, roughly double its level at the time of Meta's acquisition interest.
One-line take｜Manus is priced as the default Chinese general agent while the major platforms build their own. Being the independent option is worth a great deal, right up until it isn't.
Source: TMTPost (2026-10-13)
ByteDance launched its October 2026 share buyback at $241.35 per share pre-tax, matching the level rumored in early September, with recent departures offered the same price as current staff.
What it means for you｜It says the internal market is deep enough to absorb the exit without a haircut.
One-line take｜Paying leavers the same as employees is a liquidity signal, not generosity. It says the internal market is deep enough to absorb the exit without a haircut.
Source: ITHome (2026-10-13)
At the China International Industry Fair on October 12, Unitree founder Wang Xingxing said embodied AI could hit its ChatGPT moment in one to two years, and suggested next year's GPT-7 launch might "directly crack" the generalization problem.
One-line take｜Two months ago he said five to ten years. Moving a public timeline that fast usually means he has seen the next round of internal demos.
Source: TMTPost (2026-10-13)
UBTech ranked No. 11 globally and first among Chinese companies on Fortune's 2026 Future 120 list, compiled with BCG, with a total score of 115.
What it means for you｜UBTech being the highest-ranked Chinese name says the money is coming from deployments.
One-line take｜Growth rankings reward order books, not stage routines. UBTech being the highest-ranked Chinese name says the money is coming from deployments.
Source: QbitAI (2026-10-13)
DEEP Robotics is building its IPO story around quadruped and industrial robots that run in the field, arguing that labor substitution rather than public demonstrations is what drives volume.
What it means for you｜An IPO built on field hours instead of viral clips is the healthier kind.
One-line take｜The companies that skip the gala are usually the ones with contracts. An IPO built on field hours instead of viral clips is the healthier kind.
Source: TMTPost (2026-10-13)
Chinese robotics startup Vista Robotics closed an angel round with participation from Huawei's Hubble Investment, a Shenzhen state fund and Gaoxuan Investment, with Duowei Capital as sole adviser. The company builds flexible photonic-crystal tactile materials and is targeting embodied AI, consumer electronics and industrial scenarios.
One-line take｜Huawei's money keeps landing on components rather than platforms. That is the same playbook it ran in chips, and it tends to work.
Source: ITHome (2026-10-13)
Yinghe Robotics, founded in 2020 with backing from Midea's Yingfeng Group and later from ByteDance and Meituan, is in liquidation proceedings. The Shenzhen Intermediate People's Court accepted creditors' petition in July, and the first creditors' meeting is set for October 19.
What it means for you｜Warehouse robotics is already consolidated, and a late entrant still has to sell something to somebody.
One-line take｜Famous names on the cap table did not create a market. Warehouse robotics is already consolidated, and a late entrant still has to sell something to somebody.
Source: TMTPost (2026-10-13)
At the 2026 Paris Motor Show, BYD, Chery, Geely, Changan, Li Auto, AITO and Avatr are showing EVs and plug-in hybrids, with Chinese brands accounting for roughly 30% of the exhibits. BMW and Porsche are absent, and Volkswagen's CEO used the opening to call for European cooperation.
What it means for you｜European makers now have to defend their home show, not just their home market.
One-line take｜A third of the floor is not a guest appearance. European makers now have to defend their home show, not just their home market.
Source: QbitAI (2026-10-13)
XPeng will hold the China debut and pre-sale launch of its MONA L05, a five-seat SUV aimed at younger buyers, on October 15 at 19:00 Beijing time.
What it means for you｜XPeng moved downmarket without inventing a second badge, which few of its peers have managed.
One-line take｜MONA began as the volume line and is now the growth engine. XPeng moved downmarket without inventing a second badge, which few of its peers have managed.
Source: ITHome (2026-10-13)
With the G9L launching globally in Paris, XPeng chairman He Xiaopeng said the company is aiming for 1 million annual global sales by the end of 2030, half of it in Europe, and noted it led Chinese new-energy brands in French sales in the first half of 2026.
What it means for you｜It also puts a number on how much capacity Europe will be asked to absorb.
One-line take｜Half a million cars from a single overseas region is the most concrete China-to-Europe target posted this year. It also puts a number on how much capacity Europe will be asked to absorb.
Source: ITHome (2026-10-13)
BYD launched its Han flagship sedan from ¥235,900, with a second-generation blade battery, fast charging, Devialet audio, dual zero-gravity front seats, rear-wheel steering and the DiSus-A body-control suspension as standard.
What it means for you｜Feature lists like this are how Chinese brands keep cutting prices without much pain.
One-line take｜Rear-wheel steering on a mass-market sedan used to signal a luxury car. Feature lists like this are how Chinese brands keep cutting prices without much pain.
Source: ITHome (2026-10-13)
The second model from AIVA, the brand under Saidian Technology, a venture between Seres and ByteDance, is expected to be an SUV. Whether it adopts CATL's "chocolate" battery-swap solution is still undecided, people familiar with the plan said.
What it means for you｜This is the template for how Chinese tech firms enter cars without building them.
One-line take｜ByteDance brings software and distribution, Seres brings the chassis. This is the template for how Chinese tech firms enter cars without building them.
Source: ITHome (2026-10-13)
Hello's Robotaxi unit released HelloWorld, a generative driving world model that combines scene control, continuous generation, multi-sensor observation and efficient inference, for building driving data and testing policies interactively.
What it means for you｜Whoever generates better corner cases trains more cheaply, and that is the whole game now.
One-line take｜Every robotaxi operator is building world models because road miles are too slow to collect. Whoever generates better corner cases trains more cheaply, and that is the whole game now.
Source: ITHome (2026-10-13)
The Chinese 3D reconstruction company Qiyu Innovation has built a Gaussian-splatting model of Minning Town covering 23.5 km², trained on 186,000 aerial and ground images and 15B radar points and containing about 6.2B Gaussian ellipsoids.
What it means for you｜Physical AI needs surveyed ground truth, and someone just started selling it by the square kilometre.
One-line take｜Reconstruction at this density is a data-pipeline achievement before it is a rendering one. Physical AI needs surveyed ground truth, and someone just started selling it by the square kilometre.
Source: Leiphone (2026-10-13)
After a summer of record raises, investors in Chinese embodied-AI startups are looking to cash out early, offloading existing stakes to lock in paper gains before the cycle turns, according to founder interviews.
What it means for you｜That gap is the most useful valuation signal in the sector right now.
One-line take｜Founders are still announcing rounds while their earliest backers shop for exits. That gap is the most useful valuation signal in the sector right now.
Source: Huxiu (2026-10-13)
A supply chain has grown around motion data. Ordinary people recording repetitive tasks at home, in factories or on delivery routes sell the footage to large labs as training material for robot manipulation.
What it means for you｜The cheapness of that pipeline is a bigger competitive advantage than any single hardware design.
One-line take｜Robot hands learn from human hands, and the humans are paid by the clip. The cheapness of that pipeline is a bigger competitive advantage than any single hardware design.
Source: TMTPost (2026-10-13)
📖 The full EastVoice issue on the Chinese forum