PHYSIX FRONTIER EASTVOICE
2026.10.13 · Tuesday
Issue 059
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
490 tokens/s
Zhipu AI GLM-5.3-Highspeed output speed reported by users
¥2.3 billion
Infinigence Wuwen Xinqiong Series C funding, cumulative ¥4.3 billion
432.1%
IDC humanoid units shipped H1 year-on-year growth
$7 billion
Amazon potential valuation for AI startup Decart acquisition talks
Top Stories 3 · dual commentary ｜ Section Picks 27 · one-line takes
DeepSeek Weighs a Move Into Embodied AI, Alibaba's Qwen Ships QwenGyre, and GPT-6 Astra and an Amateur Mathematician Crack a 57-Year-Old Problem.
DeepSeek is evaluating whether to enter embodied AI, reported Huxiu, as founder Liang Wenfeng holds a flurry of meetings with robotics founders. The company is deciding between building in-house and backing outside teams.
Editor's take｜DeepSeek built its name on cheap, capable language models. Robotics is a different game — it burns cash on hardware and data collection, not just GPUs. A model lab walking in here has to relearn how to spend.
Source: Huxiu (2026-10-12)
Alibaba's Qwen team released QwenGyre, an online reinforcement-learning framework aimed at long-horizon agent tasks, according to ITHome.
Editor's take｜"Long-horizon" is the honest word here. Most agent demos fall apart after a few steps, and a framework aimed at that gap is worth more than another leaderboard entry.
Source: ITHome (2026-10-12)
A hobbyist mathematician used OpenAI's GPT-6 Astra to prove a 57-year-old geometry conjecture, QbitAI reported, the latest in a run of AI-assisted proofs.
What it means for you｜If the barrier to serious math drops this fast, the field's gatekeeping changes with it.
Editor's take｜The point is not the proof. It is that a non-professional could get there at all. If the barrier to serious math drops this fast, the field's gatekeeping changes with it.
Source: QbitAI (2026-10-12)
27 AI stories worth your time, each with a one-line take.
Huawei's Noah's Ark Lab, working with UCL, built an AI that leans on recursive self-improvement (RSI) to beat 25 games, at one point reaching a 21-0 record while calling the large model less and less as it played.
One-line take｜The headline is the win rate. The real signal is the falling reliance on the model mid-game — that is where an agent stops being a chatbot and starts being a policy.
Source: QbitAI (2026-10-12)
Microsoft released Microsoft-Decision-1, a decision model post-trained from Qwen3.5-9B, which it says beats general LLMs and rival decision models on latency and accuracy. Xbox's research team has already run it on more than 10,000 player cases.
One-line take｜A US giant quietly fine-tuning on a Chinese open model is a market signal, not a scandal. Whoever wins the open-weight base layer gets to charge rent on everyone's downstream work.
Source: Leiphone (2026-10-12)
Zhipu AI pushed out GLM-5.3-Highspeed, a new model now available in the Zcode environment, with users reporting output speeds of 490 tokens/s.
What it means for you｜When running a high-volume agent, latency stops being a benchmark and starts being a bill.
One-line take｜Raw speed is an unglamorous metric until you are running a high-volume agent. Then latency stops being a benchmark and starts being a bill.
Source: QbitAI (2026-10-12)
Intel CEO Lip-Bu Tan said Chinese frontier AI models carry low token costs and are spreading very quickly, Reuters reported.
What it means for you｜Token price is the metric that decides which models developers actually reach for.
One-line take｜When a US chip CEO says it out loud, it stops being a niche observation. Token price is the metric that decides which models developers actually reach for.
Source: ITHome (2026-10-12)
On arena.ai's week-41 ranking (2026-10-05 to 10-11), Anthropic's new Claude Opus 5.5 Max took first place on the agent board in its debut, while Chinese models held steady across the upper mid-range.
One-line take｜Chinese labs clustering in the "solid middle" is a durable position, not a defeat. They are not chasing the top of the board — they are chasing the price at which everyone else has to sell.
Source: ITHome (2026-10-12)
The Qiu Chengtong Mathematics and AI Laboratory Alliance launched at Tsinghua University, targeting dedicated math models, AI-assisted formal proofs, and interpretable AI.
One-line take｜Math is where AI's claims can be checked, which makes it the right place to build a lab. Proofs do not take a model's word for it.
Source: ITHome (2026-10-12)
BYD announced a strategic partnership with Honor to fuse AI voice across phone and vehicle, letting YOYO find and prep the car remotely and mirror the phone desktop into the cockpit.
What it means for you｜The test is whether the plumbing survives a real commute, not a launch demo.
One-line take｜Phone-and-car voice fusion has been promised for years and shipped half-built. The test is whether the plumbing survives a real commute, not a launch demo.
Source: ITHome (2026-10-12)
AMD unveiled Helion, a rack-scale solution for agentic AI workloads whose all-CPU configuration scales to 32,768 cores.
One-line take｜All-CPU racks sell one idea: not every agent step needs a GPU. If agent workloads are mostly branching logic, AMD is pricing for the part everyone else overbuilt.
Source: ITHome (2026-10-12)
AI infrastructure company Wuwen Xinqiong (Infinigence) raised ¥2.3 billion in Series C, taking cumulative funding to ¥4.3 billion — a record for China's AI-native infrastructure sector.
One-line take｜Infrastructure funding records are a market bet, not a product milestone. The money says compute is still scarce; the runway says they have to sell it before that stops being true.
Source: ITHome (2026-10-12)
Tencent's WorkBuddy workspace added StepFun's flagship Step 5 Preview, free to users from October 12 to 25.
One-line take｜A two-week exclusive giveaway is a distribution play, not a pricing one. StepFun is buying user exposure with a discount Tencent can afford to hand out.
Source: ITHome (2026-10-12)
Tencent's Yuanbao team is developing "Yuanbao Assistant," a Personal Agent product planned as a standalone app with its own runtime and permissions for long-cycle tasks.
One-line take｜A separate app with separate permissions tells you Tencent does not fully trust the main app's interface for long-running work. That is a candid read on the limits of chat.
Source: ITHome (2026-10-12)
Sportswear group Anta rolled out agents into content production, overseas ordering, and cross-border returns, cutting order placement to 10 minutes and building over 1,000 orders a month.
What it means for you｜Cutting order cycles to minutes is the kind of number that gets copied, quietly, across the sector.
One-line take｜Anta is not showcasing a chatbot, it is publishing a labor line. Cutting order cycles to minutes is the kind of number that gets copied, quietly, across the sector.
Source: Leiphone (2026-10-12)
Amazon is negotiating to buy San Francisco-based Decart at a potential valuation of about $7 billion, per the Wall Street Journal. Decart builds cross-chip portability software.
One-line take｜A $7B price for portability software says the real fear is hardware lock-in. Buying the layer that frees models from any single chip is a bet against everyone's GPU strategy, including its own.
Source: ITHome (2026-10-12)
Kling AI, Kuaishou's AI video-generation product launched in June 2024, has been tagged by investors at roughly ¥100 billion despite not yet being profitable, monetizing through subscriptions and enterprise APIs.
What it means for you｜Video generation has to become a workflow before that math closes.
One-line take｜A nine-figure valuation on a product that does not yet make money is a wager on format, not on revenue. Video generation has to become a workflow before that math closes.
Source: Huxiu (2026-10-12)
Mosi Intelligence closed a Pre-A round of more than ¥1 billion, betting on a next-generation all-modality base model for real-time interaction, reported AI Era.
One-line take｜Real-time multimodal interaction is the demo everyone shows and few can serve at scale. A ¥1B seed says investors expect someone to fix the serving problem soon.
Source: AI Era (Xinzhiyuan) (2026-10-12)
At the Paris Motor Show, XPeng said it will reach mass production of its humanoid robot by the end of 2026, with deliveries starting in the first half of 2027.
One-line take｜"Mass production" and "deliveries six months later" tell two different stories. A carmaker's real edge here is its supply chain, not the robot itself.
Source: ITHome (2026-10-12)
At the CIIF expo opening, Unitree founder Wang Xingxing said his biggest mistake over the past decade was not believing in AI strongly enough.
One-line take｜Coming from the man who sells China's most visible humanoid, this is a repositioning more than a confession. Unitree is announcing that the body is not the moat.
Source: ITHome (2026-10-12)
TMTPost reported that IDC counted nearly 25,000 humanoid units shipped globally in H1, up 432.1% year on year, while Sullivan put the figure at 36,000. Revenue-share leader UBTECH still faces a different set of books.
What it means for you｜When the market size is that fuzzy, 'who earns' is closer to a marketing question than an accounting one.
One-line take｜A 40% gap between two research houses on the same shipment number is the story. When the market size is that fuzzy, "who earns" is closer to a marketing question than an accounting one.
Source: TMTPost (2026-10-12)
QbitAI reported that robotics and physical-AI companies have raised close to $48 billion (~¥321 billion) in 2026 so far, averaging about ¥1 billion a day.
What it means for you｜The money is now big enough that a few visible failures will move the whole flow.
One-line take｜A billion yuan a day into embodied AI is not a seed-stage trend, it is an asset class forming. The money is now big enough that a few visible failures will move the whole flow.
Source: QbitAI (2026-10-12)
CRRC Changchun and Chengdu Metro unveiled VELLINK, billed as China's first "embodied-intelligence" digital metro train, able to optimize operations and set running strategies on its own.
One-line take｜Putting "embodied" on a metro train stretches the word, but the real work — predictive operations on fixed infrastructure — is where AI quietly earns its keep in transit.
Source: ITHome (2026-10-12)
China's passenger-car retail sales fell 28.0% year on year in September to 1.618 million units, per the CPCA, while new-energy vehicles reached a record 67% penetration.
What it means for you｜The pie is shrinking and being handed to electric players at the same time.
One-line take｜A sharp overall drop paired with a record NEV share is the whole story of this market. The pie is shrinking and being handed to electric players at the same time.
Source: ITHome (2026-10-12)
After the Zunjie V800 brake-pedal bracket failure went viral, the brand rolled out a remedy, but some deposit-holders said they would walk away even without a refund.
What it means for you｜Buyers calling a ¥50,000 deposit disposable is the costliest signal a new brand can receive.
One-line take｜A fix can be issued in days; trust comes back in quarters. Buyers calling a ¥50,000 deposit disposable is the costliest signal a new brand can receive.
Source: Huxiu (2026-10-12)
Tesla said the Model Y was the best-selling passenger car of any type in Denmark, Finland, Iceland, Norway, and Sweden in September, and that one in four new cars in Norway was a Tesla.
What it means for you｜The Nordics are the clearest preview of a ceiling every EV maker will eventually hit.
One-line take｜One in four new cars in Norway is a saturation number, not a growth one. The Nordics are the clearest preview of a ceiling every EV maker will eventually hit.
Source: ITHome (2026-10-12)
Chinese automakers used the Paris Motor Show to push into the premium tier, with brands like XPeng and Li Auto rolling out flagship models and European launch plans while German incumbents scaled back their presence.
One-line take｜A century-old European show turning into a Chinese product stage is a shift in who sets the terms. The cars are the message; the venue is the proof.
Source: Huxiu (2026-10-12)
Wang Pengwei, who leads BAAI's embodied-model research center, argued in TMTPost that current world models capture only a fragmented picture of reality and are not yet a path to general physical intelligence.
One-line take｜World models are the field's favorite promise right now, so an insider calling them partial is worth weighing. The gap he names is data, not architecture.
Source: TMTPost (2026-10-12)
QbitAI reported that ByteDance's latest generative system turns a single sentence into an explorable, interactive miniature world, moving beyond image generation into scene generation.
What it means for you｜For games and simulation, that is a much larger shift in who gets to build.
One-line take｜Image generation changed who could make a picture. Interactive worlds change who can make a place. For games and simulation, that is a much larger shift in who gets to build.
Source: QbitAI (2026-10-12)
Investor Wang Yuquan, in a Huxiu interview, predicted the global AI market may split roughly 7-to-3 in China's favor and argued that embodied AI is AI's next stop, though expectations are running ahead of reality.
One-line take｜A 7-to-3 call is a forecast you can hold someone to. He pairs it with a warning against over-excitement on embodied AI — the two together are the more honest read.
Source: Huxiu (2026-10-12)
📖 The full EastVoice issue on the Chinese forum