China Desk — 26 Aug 2026

Zhipu reveals GLM-5.3-Flash, built for multimodal work on domestic chips

# QbitAI

Chinese coverage stresses that the formerly anonymous Ox Alpha model is Zhipu’s open 320B GLM-5.3-Flash, matching a leading Claude score while serving 62T tokens on domestic chips.

GLM-5.3-Flash is the first native multimodal model in the reported GLM-5 line, accepting images, video, files and text rather than bolting vision onto a text-only path.
The reported architecture combines sparse and linear attention to reduce long-context cost, while the 6B active-parameter claim shows where Chinese model makers are seeking efficiency rather than only larger totals.
The domestic-chip deployment is the strategic headline: Chinese coverage presents high-volume overseas use on local hardware as evidence that the ecosystem can support frontier workloads under export constraints.
The article reports parity with Claude Opus 4.8 and prices as low as 0.8 yuan ($0.11) input and 2.8 yuan ($0.39) output per million tokens, but the comparisons come from company or publisher tests.
The anonymous test launch also reveals a product strategy: gather real traffic and skepticism before attaching the domestic brand, then convert that attention into an open-model release.

Open at QbitAI Ask Claude ChatGPT

China’s robot competitions start measuring whether machines can hold a job

# Zhidx

At a Chinese humanoid-robot competition, Zhiyuan won 18 gold, 16 silver and 12 bronze medals as the event expanded autonomous tasks in factories, libraries, logistics and emergency response.

The event shifted from spectacle toward work simulation: scenario events rose from six to 21, and autonomous completion received full weight while remote control received half weight.
Zhiyuan’s emergency and library systems combine a vision-language model, an agent for task decomposition and reinforcement-learning tools for policy training and validation.
The company reports that its G2 robots already ran tablet quality-control tasks for six days and processed more than 17,500 units at a 99.99% success rate, though those are company-provided figures.
The competition is becoming an industrial benchmark because unfamiliar environments and limited preparation test generalization, while factory deployments test repeatability and uptime.
The strategic advantage claimed by Chinese robot makers is a closed loop between body, model and real deployment data; the remaining question is whether such performance transfers beyond curated event scenarios.

Open at Zhidx Ask Claude ChatGPT

Alibaba open-sources Qwen3.8-Flash with a 6B active core

# Leiphone

Alibaba released Qwen3.8-Flash as a sparse multimodal model with 125B Transformer parameters but only 6B active per token, claiming major gains in agent benchmarks and serving efficiency.

Also: Hacker News

The architecture combines Qwen Sparse Attention with gated linear attention, a gated residual design and a large N-gram embedding table that acts as a low-cost memory extension outside the main Transformer path.
Alibaba says the model was trained with about one-ninth the resources of Qwen3.7-Plus and costs 1 yuan ($0.14) per million input tokens and 3 yuan ($0.42) per million output tokens before discounts.
The model’s reported strengths are agent coding, long professional tasks and multimodal device interaction, suggesting that sparse activation is being optimized around workflow demand rather than only static benchmark scores.
The release is also a platform move: Alibaba put the model into Qwen Office, exposed it through API services and open-sourced weights for community testing as a possible foundation for Qwen4.
The benchmark claims come from Alibaba and associated coverage, so the durable signal is the design and pricing strategy, not the unverified claim of global superiority.
Chinese model competition is increasingly about intelligence per watt and per yuan, with architecture, post-training and application harnesses being designed together.

Hacker News debates Qwen3.8-Flash’s real business advantage Hacker News

Open at Leiphone Ask Claude ChatGPT

China’s week: border disaster, summit diplomacy and a possible Politburo meeting

# Sinocism

Chinese-focused newsletter coverage connects the Gyirong disaster with Xi Jinping’s upcoming regional summit travel and the expected August Politburo meeting, where economic support may be tested.

The newsletter says Xi is expected to attend the Shanghai Cooperation Organisation summit and visit Egypt, placing foreign travel alongside a potentially consequential domestic policy meeting.
Its forecast that the Politburo meeting could produce little new economic support is explicitly an expectation, not a reported decision, but it captures current policy uncertainty.
The disaster section and the political calendar sit together because Chinese governance treats emergency response, diplomacy and economic management as one centralized agenda.
The item is valuable as a map of what China-watchers are watching before official documents appear, rather than as a single confirmed event.

Open at Sinocism Ask Claude ChatGPT

China makes officials liable for serious environmental damage for life

# SCMP

New Party and State Council rules tie serious environmental harm to lifelong official accountability, strengthening the link between ecological performance and cadre evaluation.

The revised rules apply to officials even after they leave a post, according to the supplied report, extending environmental responsibility beyond the normal administrative tenure.
The measure fits China's broader shift from pollution campaigns toward a governance model in which local officials are judged on the quality and durability of growth.
Linking environmental outcomes to personnel discipline is more consequential than another emissions target because it changes incentives inside the state hierarchy.
The report does not specify which agencies will investigate cases, how causality will be established, or how often the rule will be enforced.
The policy could improve compliance where local governments previously traded environmental damage for investment, but it may also encourage concealment or conservative decision-making if metrics are unclear.
For industrial companies, the strategic implication is a longer tail of environmental liability and greater importance of documentation, monitoring, and remediation.
The rule matters most if it produces high-level disciplinary cases or changes how local investment projects are approved.

Open at SCMP Ask Claude ChatGPT

SenseTime reports its first IFRS profit as token production scales

# Leiphone

SenseTime says it reached a $87M first-half IFRS profit while its token factory served 2.4 trillion tokens per day and its agents reached more than 1,000 enterprise customers.

The company’s three-part strategy is one model, one token factory and one agent-control system, linking model quality, serving cost and task delivery rather than treating them as separate businesses.
SenseTime reports 4.8 trillion? The supplied source states 4.8 wan P of operating compute, while daily service reached 2.4 trillion tokens; the figures indicate scale but need careful interpretation because the units are not directly comparable.
Its domestic-chip work reportedly lifted model-utilization efficiency by up to 2.5 times over manufacturer baselines and achieved early compatibility across more than ten Chinese chip vendors.
The company says its agents serve government, finance, education, logistics and other sectors, while consumer agents reached 45 million users, giving it both enterprise deployment and mass-market feedback.
The first profit was helped by fair-value gains from ecosystem assets, so recurring operating economics still need to be separated from accounting revaluation.
The strategic test is whether SenseTime can shift from selling compute or access toward pricing completed tasks while maintaining margins and reliability.

Open at Leiphone Ask Claude ChatGPT

Guming lets an agent triage database incidents before humans intervene

# GeekPark

Chinese tea chain Guming uses an agent to diagnose database alerts, recommend safe actions and route approved changes through a gateway, cutting DDL approval time 30% to 40%.

The deployment starts with constrained, repeatable work: metadata appears inside the developer workflow, low-risk changes can be executed automatically and high-risk actions remain human-approved.
For incidents, the agent pulls slow logs, compares historical baselines and offers options such as indexes, throttling or temporary scaling instead of directly improvising on production.
A data gateway parses and rewrites SQL, masks unauthorized fields and records activity by session, tool and scenario, turning agent permissions into an auditable boundary.
The organizational result is as important as the latency reduction: DBAs spend less time on alert handling and more on capacity planning, engine upgrades and architecture.
This is a useful Chinese enterprise case because it shows the adoption sequence that works in practice: standardize the workflow, constrain execution, collect feedback and only then widen autonomy.

Open at GeekPark Ask Claude ChatGPT

China’s internet giants move their AI battle into office software

# Leiphone

Tencent, Alibaba and ByteDance are redirecting AI resources toward enterprise office tools after consumer chatbot campaigns failed to convert attention into durable revenue.

Chinese coverage frames the shift as a second commercialization attempt: consumer acquisition created users, while office workflows offer clearer buyers, budgets and recurring usage.
Tencent is pushing WorkBuddy, Alibaba is combining QoderWork and other tools inside Qwen Office, and ByteDance is aligning Doubao with Feishu and Volcano Engine.
Coding was the first enterprise beachhead, but the domestic market remains small; office software offers a much larger installed base and more opportunities to embed AI into daily work.
The strategic contest is not simply model quality. It is distribution, workflow integration, enterprise support and the ability to turn habitual use into continued payment.
A new Chinese rule governing anthropomorphic interactive services may also help companies reposition agents from entertainment products toward bounded productivity systems.

Open at Leiphone Ask Claude ChatGPT

Xiaomi unveils three in-house AI chips across phones and cars

# 36Kr

Xiaomi launched phone, edge-AI and autonomous-driving chips, extending a self-designed silicon strategy across its ecosystem after more than 21 billion yuan, roughly $3.0 billion, in reported investment.

Also: Leiphone

The O3 phone SoC uses a 3 nm process and 24 billion transistors, while the O100 combines an NPU with stacked memory to attack the bandwidth wall that increasingly limits edge inference.
The O100's reported 1.22 TB/s internal bandwidth and up to 330 tokens per second reflect a near-memory-compute design using two DRAM wafers bonded to an NPU wafer.
The D100 targets driving and personal AI compute with 20 CPU cores, a 16-core NPU, up to 160 GB of memory and support for local models above 200B parameters, according to Xiaomi's presentation.
Multi-chip fusion is intended to make the driving platform expandable rather than locking a vehicle's intelligence ceiling to one accelerator generation.
Xiaomi says it has invested more than 18 billion yuan, roughly $2.5 billion, over five years and has a 50 billion yuan, roughly $7.0 billion, total budget, with a team above 3,000 people.
The chips are primarily for Xiaomi's own products, trading external chip sales for tighter hardware, software and vehicle co-design; that makes the strategy powerful but leaves the company carrying the full capital burden.
The important China signal is that edge AI is pulling consumer-electronics companies deeper into semiconductor design, especially where privacy, latency and local-model capacity are product differentiators.

Xiaomi's chip strategy is finally becoming an ecosystem Leiphone

Open at 36Kr Ask Claude ChatGPT

US sanctions Chinese entities over trade with Iran

# SCMP

The US Treasury has targeted about 24 mainland Chinese or Hong Kong entities in an Iran-related sanctions campaign, while Chinese officials say cooperation with Iran follows international law.

The action widens the compliance perimeter for Chinese firms that facilitate Iranian trade, even when they are not directly involved in Iran’s government or military.
Beijing’s response is framed around sovereign rights and legal trade rather than a defense of the named companies, a familiar pattern in sanctions disputes.
For multinational companies, the practical risk is secondary exposure: banks, logistics providers and technology suppliers may avoid counterparties touched by the designation even where the commercial transaction is lawful in China.
The episode sits inside a larger contest over whether Washington can enforce economic pressure through global financial networks while Beijing builds alternative trading relationships.
The sanctions will matter more if they change the behavior of Chinese intermediaries or trigger broader restrictions; the excerpt does not indicate either outcome yet.

Open at SCMP Ask Claude ChatGPT

China threatens retaliation over expanded US Iran sanctions

# SCMP

Beijing says it may retaliate against new US sanctions on Chinese entities tied to Iran, creating a fresh test of bilateral stability before expected Xi-Trump engagement.

The dispute extends US-China friction beyond technology and trade into secondary sanctions and energy relationships, where Beijing objects to Washington applying its rules to Chinese firms.
Chinese retaliation could target entities or sectors without necessarily triggering a broad rupture, giving both sides room to signal resolve while preserving a wider diplomatic channel.
The timing matters: a new sanctions flashpoint can complicate summit preparation even when both governments publicly emphasize stability.
The source presents observer assessments and does not specify the final Chinese measures, so the immediate signal is threat formation, not confirmed escalation.

Open at SCMP Ask Claude ChatGPT

DeepSeek open-sources a modular runtime for autonomous agents

# InfoQ China

DeepSeek released an MIT-licensed Harness preview that splits model adapters, tools, sandboxes, state and event logging into replaceable runtime plugins.

The microkernel design separates the agent loop from model endpoints, tool registries, execution environments and user interfaces, allowing a developer to swap components through declarative configuration.
An append-only event log records messages, tool calls, intermediate states, token metrics and sub-agent dispatches, making replay and failure isolation first-class runtime functions.
This is a meaningful architectural signal from China: the model lab is treating agent execution infrastructure as an independent open-source layer rather than hiding it inside a monolithic product.
The preview supports several runtime profiles, but its long-term value will depend on stable extension contracts, compatibility and adoption beyond DeepSeek’s own ecosystem.

Open at InfoQ China Ask Claude ChatGPT

China and India agree on eight steps to ease border tensions

# Jiemian

Chinese coverage presents the eighth special-representative meeting as a concrete de-escalation package, including expert groups, new military hotlines, trade, pilgrimages, and river talks.

China and India agreed to create groups on boundary delineation and border management under their existing consultation mechanism.
The two sides will add two general-level meeting points and two military communication hotlines in the eastern and central sectors.
The package keeps the border dispute inside a managed negotiation framework rather than claiming a final settlement.
It also links security stabilization to reopened border markets, pilgrimage access, and a future meeting on transboundary rivers.
The agreement schedules the next special-representative meeting in India in 2027, while the supplied account does not establish how quickly the new mechanisms will operate.
Its importance depends on whether the new channels prevent local incidents from escalating and whether political leaders convert procedural contact into substantive boundary talks.

Open at Jiemian Ask Claude ChatGPT

China establishes a two-way laser link across the Earth-Moon distance

# IT Home

Chinese researchers report a two-way laser link across more than 400,000 km, moving space optical communications from near-Earth trials into lunar-distance engineering.

The test achieved an initial 1.25 Mbps uplink and 100 Mbps downlink, according to the Chinese Academy of Sciences source cited by the report.
The receiver had to recover signals reduced to only a few photons amid moonlight, starlight, and urban background noise.
The pointing system modeled orbital motion, telescope error, atmospheric refraction, and signal travel time so both ends could track each other dynamically.
A single-photon detector and signal-recognition algorithms addressed the link-budget and noise problem rather than merely increasing transmitter power.
The reported example cuts transmission of an 8K lunar image from roughly four or five minutes over a 5 Mbps microwave link to about 12 seconds.
This is an enabling communications layer for higher-data lunar science and future surface operations, though the source describes an experiment rather than an operational network.

Open at IT Home Ask Claude ChatGPT

Earthquake-warning dispute exposes a fight over authority and reliability

# IT Home

A Chengdu disaster institute apologized for a magnitude error but rejected the label of false alarm, while the earthquake bureau accused it of unauthorized nationwide alerts.

The private institute's first alert for a magnitude 4.7 Sichuan earthquake reported magnitude 7.7, a three-magnitude error that it says lasted one second.
The institute argues that an alert during an unfolding earthquake is inherently provisional and that a false alarm technically means an alert without an earthquake.
The provincial earthquake bureau says the institute used its own system while distributing the alert under the China Earthquake Early Warning Network name, after third-party authorization had ended.
This is therefore a governance conflict as much as a sensor-accuracy dispute: two institutions are contesting branding, authorization and responsibility for public safety.
The institute says its subscription service has accumulated roughly 100 million users, creating a political argument over who bears responsibility if its alerts are blocked.
The episode exposes a missing institutional layer in China's warning system: common evaluation standards for average error, worst-case error, response time and blind-zone size.
The practical test is whether regulators produce shared technical and accountability rules, rather than allowing the dispute to remain a contest between state authority and an established private network.

Open at IT Home Ask Claude ChatGPT

China's new-energy vehicle fleet passes the 10 million scale

# IT Home

China's industry ministry says new-energy vehicles rose from 1.367 million annual sales to 16.49 million during the last five-year period, reaching 47.9% of new-car sales.

The ministry says new-energy vehicles, batteries and key materials account for more than 70% of global output and Chinese products reach over 100 countries.
July new-energy passenger-car penetration reached a record 65.1%, even as overall passenger-car retail sales fell sharply year on year.
The next five-year plan will focus on product quality, automated-driving safety, intelligent-vehicle access trials and vehicle-road-cloud pilots.
The policy language pairs industrial expansion with tighter governance: production-consistency checks, consumer protection and clearer rules for automated driving.
For manufacturers, the strategic shift is from proving demand to managing safety, quality and international standards at scale.

Open at IT Home Ask Claude ChatGPT

Chinese rocket stage completes two-year orbital-service mission

# IT Home

Galactic Energy says its repurposed rocket-stage platform completed two years of tests and a controlled deorbit, validating a full commercial in-orbit service loop.

The platform converted a launch vehicle’s spent upper stage into a long-duration service vehicle by upgrading thermal control, power and attitude-control systems.
The reported sequence covered orbital insertion, configuration change, relocation from 535 to 580 kilometers, payload testing and precision deorbiting.
During its mission it tested millinewton electric propulsion and other commercial payloads in a real space environment, providing engineering data that ground tests cannot reproduce.
The result is strategically significant because it turns otherwise discarded launch hardware into a reusable orbital test and service asset.
A second platform is already in orbit, and the company says it plans a dedicated orbital compute demonstrator in 2027.
The next proof point is economics: repeated customer missions and reliable end-of-life disposal must make the architecture cheaper and more useful than dedicated small satellites.

Open at IT Home Ask Claude ChatGPT

China says HarmonyOS is now the world’s third-largest mobile OS

# IT Home

China’s industry ministry says HarmonyOS has become the third-largest mobile operating system and that open-source HarmonyOS devices exceed 1.35 billion, framing software independence as industrial infrastructure.

The figures were presented at a State Council Information Office briefing, so they are government claims rather than an independently defined global market-share measurement.
The ministry also reported more than 1.35 billion devices in the open-source HarmonyOS ecosystem, spanning phones, cars, appliances and industrial systems; the number is broader than smartphones alone.
The announcement places operating systems alongside databases, industrial software and artificial-intelligence open-source communities in the country’s industrialization agenda.
Huawei’s consumer HarmonyOS milestone and the broader open-source ecosystem are related but not identical measures, an important distinction when assessing adoption.
The strategic goal is ecosystem depth: a domestic software layer that can serve many device categories and reduce dependence on foreign operating-system platforms.
Whether HarmonyOS can become a durable global alternative depends on developer tooling, application availability and adoption outside China, none of which the announcement establishes.

Open at IT Home Ask Claude ChatGPT

Ant International pushes time-series models into financial operations

# Signal QbitAI

Ant International says its FalconTST 2.0 reached a 0.666 MASE score and more than 93% forecast accuracy, with banks using it for cash-flow and foreign-exchange risk work.

The model’s significance is specialization: it starts with a general time-series foundation and adds domain validation in low-signal, heavy-tailed financial data rather than assuming a generic benchmark transfers directly.
FalconTST’s evolution separates efficient single-series prediction from heterogeneous multivariate modeling, reflecting the different structures found in payments, foreign exchange and supply chains.
The article names Barclays, Citi, Deutsche Bank and Standard Chartered as users, but the performance and adoption claims come from company reporting and do not show trading profit.
The broader research context is a warning against treating time-series foundation models as automatic alpha: domain retraining, rolling backtests and risk constraints remain necessary.
This is a useful Chinese example of AI commercialization in finance because the product is sold as a validated workflow component, not simply as a downloadable model.

Open at QbitAI Ask Claude ChatGPT

DeepSeek’s revenue jumps tenfold while infrastructure spending explodes

# Signal Zhidx

Reported figures put DeepSeek’s January-to-July revenue at $67M and loss at $101M, alongside $1.55B of infrastructure spending and a proposed $7B financing round.

The reported 44.6% gross margin, and 82.9% margin on API calls, suggest that serving efficiency is currently a stronger business than the company’s absolute revenue scale.
Infrastructure spending rose from about $169M in 2025 to $1.55B in the first seven months of 2026, showing how quickly a low-price model can turn popularity into capital demand.
The proposed financing and $70B valuation are attributed to people familiar with the matter, not DeepSeek disclosures, so they should not be treated as settled corporate facts.
DeepSeek’s pricing changes are part of the same commercialization push: peak rates rose sharply while off-peak rates preserved a low-cost path for less time-sensitive workloads.
The strategic question for Chinese frontier labs is whether high gross margin can compound into enough revenue to fund model development without relying indefinitely on external capital.

Open at Zhidx Ask Claude ChatGPT

Meitu bets that vertical AI video workflows beat generic models

# AI QbitAI

Chinese coverage argues that Meitu’s specialized video products win by packaging models into repeatable workflows for marketing, talking-head video and music production.

Also: Zhidx

The application thesis is that users buy a dependable result, not access to a particular model, so the winning layer can be the workflow, asset management and quality control around generation.
Meitu’s advantage is accumulated domain knowledge from years of image editing, plus user behavior that reveals aesthetic preferences and desired outcomes for future model tuning.
Products such as Kaipai, Vmake Labs and MVLAND focus on high-value commercial tasks, where a good output can generate revenue and make users less price-sensitive.
The broader Chinese market signal is a shift from model-versus-model comparisons toward specialized agents, canvases and skills that turn generation into an operating process.

Meitu’s AI productivity apps drive a profitable first half Zhidx

Open at QbitAI Ask Claude ChatGPT

Doubao Work turns Feishu's organizational context into an agent

# AI QbitAI

ByteDance's Doubao Work is deeply integrated with Feishu, letting an agent read company chats, documents and permissions instead of starting every task with an empty context.

The product combines browser operation, file handling, web research and interactive-page generation, but its differentiator is access to the organization's existing conversations and workflow state.
In a reported procurement test, it marked unavailable bulk prices, stock and delivery dates as estimates or items needing quotation rather than inventing them, a useful behavior for enterprise deployment.
Feishu login puts the agent inside the organization's identity and permission system, avoiding the separate applications and credentials commonly required by external agents.
The product can turn discussions into structured task lists, assign owners and sync results into tables, showing why enterprise context is more valuable than another generic chat interface.
This is the same underlying ByteDance event as the wider office-AI strategy: the company is trying to convert a consumer model and collaboration platform into recurring enterprise infrastructure.
The open question is governance. Reading chats, documents and approvals can make an agent useful, but it also creates a concentrated permission and audit surface for the employer.

Open at QbitAI Ask Claude ChatGPT

Unitree and Zhiyuan robots reportedly share one general-purpose brain

# AI QbitAI

A Chinese report highlights an uncut ten-minute demo in which Unitree and Zhiyuan robots perform household tasks, recover from interruptions and cooperate across different bodies.

The claimed demonstration is technically interesting because two robots with different hardware, sensors and degrees of freedom reportedly use one model and coordinate without scripted handoffs.
The tasks include force-sensitive window cleaning, object placement and interruption recovery, which test long-horizon state management rather than one isolated motion.
Cross-body cooperation is the strongest claim: the robots allegedly infer complementary roles and exchange objects without explicit verbal instructions.
The source is highly enthusiastic and provides no independent test protocol, deployment logs or failure rate, so the video should be treated as an impressive demonstration rather than proof of general autonomy.

Open at QbitAI Ask Claude ChatGPT

Huawei storage gains ground as AI changes enterprise buying criteria

# Analysis Zhidx

Chinese coverage says Huawei entered Gartner’s storage-leader group with a stronger showing in AI and high-performance file workloads, challenging long-established North American vendors.

The strategic shift is from storage capacity toward AI throughput, ransomware resilience, energy cost and supply-chain control as enterprises accumulate unstructured data.
Huawei’s Dorado system uses active-active NAS, separated control and data paths and claimed seven-nines reliability; these are product assertions, not independent market-share proof.
The Gartner position is meaningful as external validation but does not by itself show that Huawei has displaced NetApp, Dell or IBM in installed base or global revenue.
Chinese coverage frames the result as a decade-long technology and ecosystem catch-up, reflecting the domestic narrative of substituting for foreign infrastructure leaders.
The broader market question is whether AI-specific storage architectures create a durable opening for Huawei or whether incumbent vendors quickly replicate the same features.

Open at Zhidx Ask Claude ChatGPT

Chinese podcast listeners shift from AI capability to AI collaboration

# Culture QbitAI

Xiaoyuzhou’s 2026 AI report says searches for collaborative AI rose 227% and discussions increasingly ask how people should work with systems rather than what models can generate.

The report says AI-related creators grew 187% and programs 239% in 2025, making Chinese podcasting a record of how technical ideas move into work and culture.
In 2026, collaborative-AI searches rose 227%, episode count 290% and plays 207%, while workplace-AI searches rose 105.4%, showing a move from novelty toward practical use.
Discussion of human value generated 335,000 comments and 12.86 million words, with listeners focusing on judgment, taste, experience and responsibility as tasks become easier to delegate.
The platform’s audience is unusually young and practitioner-heavy, so the numbers are better read as a map of discourse and attention than as a national adoption survey.
The cultural signal is important: Chinese AI conversation is moving from model capability to division of labor, authorship and the social meaning of doing work oneself.

Open at QbitAI Ask Claude ChatGPT

OpenAI's first inference chip reportedly beats Blackwell on efficiency

# AI QbitAI

Chinese coverage highlights SemiAnalysis testing that found OpenAI's 700-watt inference chip ahead of Nvidia Blackwell on selected latency and energy metrics, with important benchmark caveats.

The supplied report claims up to 1.5 to 1.9 times the inference throughput per kilowatt of GB200 and GB300 rack systems, while using single-token prediction rather than speculative decoding.
It says the tested workloads included open models from OpenAI, DeepSeek and Moonshot AI, with low-concurrency output around 1,400 tokens per user per second for some models.
The comparison is not apples-to-apples: Blackwell results used multi-token prediction, the complete test suite was not run and the more production-like AgentX workload was omitted.
The chip reportedly uses six HBM4 stacks totaling 216GB and 15.4 terabytes per second of bandwidth, while a second revision is claimed to improve performance per watt by about 25%.
The strategic signal is that frontier model companies may design inference silicon around their own serving patterns, but engineering samples and selective benchmarks are far from displacing a mature software ecosystem.

Open at QbitAI Ask Claude ChatGPT

FaceWall puts edge intelligence at the center of its next phase

# Analysis Zhidx

Chinese model company FaceWall says its next phase will focus on efficient edge models, scaling vertical deployments and using AI inside its own organization.

The company’s density-law thesis argues that future intelligence will be measured by useful capability per watt, potentially placing capable agents inside phones, cars and industrial devices.
FaceWall presents local processing as a privacy and resilience advantage: a device can keep data on board and continue making decisions when connectivity is weak.
Its three priorities combine technology, commercialization and internal operations, suggesting that the company wants its own organization to serve as a testbed for AI-mediated work.
The strategy fits China’s manufacturing base and large device market, but the supplied letter is an internal narrative rather than evidence of broad deployment economics.

Open at Zhidx Ask Claude ChatGPT

RoboHarness lets specialized robot policies cooperate

# AI Leiphone

A Huawei research team proposes RoboHarness, a coding-agent layer that routes among VLA, world-action, reinforcement-learning and task-motion-planning policies without retraining them.

The system treats existing policies as agentic skills and adds an organizational layer that decides who should act, when to reset and how to recover after failure.
Understanding skills turn raw perception into routing signals such as training-distribution similarity, pose stability and image quality; memory skills bridge state mismatches; evolution skills update routing from feedback.
This architecture addresses a different problem from making one policy robust: it coordinates several specialists when no single model covers semantic interpretation, geometric planning, control and long-horizon reasoning.
The design is technically credible because it avoids forcing different policies to share model structure, action spaces or training data, reducing the cost of integrating heterogeneous research systems.
The open question is latency and failure arbitration: a high-level coding agent must make routing decisions quickly and safely enough for physical work.

Open at Leiphone Ask Claude ChatGPT

LiveEdit brings causal streaming video editing to 12.66 FPS

# AI Leiphone

Researchers from Tsinghua and Hong Kong University of Science and Technology propose a causal streaming diffusion system that edits video in four steps while preserving unedited regions.

LiveEdit converts a bidirectional video editor into a chunk-wise causal model, allowing it to process incoming frames without waiting for the full future sequence.
Its three-stage distillation first learns high-quality bidirectional editing, then transfers the behavior to causal chunks and finally compresses sampling from 100 steps to four.
An autoregressive mask cache reuses self-attention features for roughly 70% of redundant spatial tokens while retaining fuller computation in regions likely to change.
The design choice is technically important: reusing self-attention is safer than reusing the whole block because feed-forward layers retain more high-frequency texture and structure.
The reported 12.66 FPS result is relevant to live effects, video meetings and augmented reality, where consistency and latency matter more than one polished offline clip.

Open at Leiphone Ask Claude ChatGPT

Action-aware routing makes VLA inference 1.79 times faster

# AI Leiphone

A Chinese research team proposes AC2-VLA, which uses the robot’s action context to decide when to reuse cognition, prune visual tokens or skip layers.

The paper challenges the common assumption that visual complexity determines compute need: a simple-looking frame may require maximum precision during contact, while a visually busy transit phase may not.
A single router combines cognition-cache reuse, action-conditioned token pruning and layer skipping, allowing different computation depths at different stages of a manipulation task.
Self-distillation uses a dense policy as teacher because task success is too sparse and delayed to supervise every routing decision directly.
The method is designed to transfer across VLA backbones through renewed distillation rather than task-specific engineering, which is more useful for a fast-changing robotics stack.
The reported speedup matters operationally because a robot that pauses for seconds between control actions can be uneconomic even when its success rate is high.

Open at Leiphone Ask Claude ChatGPT

A Chinese robotics startup brings a full-stack system to the table

# AI QbitAI

Embodied-AI startup SuperDynamix displayed a robot, dexterous hand, data-capture headset, world model and training infrastructure around a table-tennis demo.

The company argues that embodied AI needs a closed loop from human demonstrations and simulated worlds to hardware execution and back into training, making a full stack more than a vertical-expansion story.
Its KAI Bot has 117 degrees of freedom and about 18,000 tactile points, while the KAI Hand combines active, passive and compliant joints for manipulation.
The table-tennis system is intended to test continuous perception, trajectory prediction and rapid control rather than a pre-recorded sequence.
The technical claims are company supplied and the startup is only about a year old, so the important signal is strategic direction rather than proven scale or reliability.

Open at QbitAI Ask Claude ChatGPT

Flash floods hit the China-Nepal border, with mass casualties reported

# Analysis SCMP

English coverage highlights dozens dead and hundreds missing after floods and landslides crossed from Nepal into China's Gyirong border area, with many missing people foreign tourists.

The disaster affected a border checkpoint, roads, bridges, electricity, and communications on both sides of the Himalayas.
Chinese state media described major casualties and missing people in Gyirong, but the supplied account did not yet provide China's death toll.
The event combines a climate hazard with the operational fragility of a remote border crossing and tourist route.
The English account adds urgency around foreign travelers and the cross-border rescue problem, while the Chinese account foregrounds the official emergency response and local fund transfers.
The figures are provisional, and the decisive unknowns are access to isolated areas, the number of Chinese casualties, and the speed of restoring transport and communications.
For infrastructure planners, the incident is a reminder that a checkpoint is a coupled system: roads, power, communications, tourism, and customs can fail together.

Open at SCMP Ask Claude ChatGPT

TokenRhythm bets on routing agents across multiple models

# AI Leiphone

Chinese AI infrastructure company TokenRhythm raised several tens of millions of dollars to build a routing harness that chooses models by task stage, cost and capability.

The company’s thesis is that model abundance turns orchestration into the scarce layer: an agent should switch models as a task moves from planning to execution or verification.
Its routing layer is intended to collect model choices, outputs and user feedback as a dataset for an eventual agent-native model, making the control plane a source of training signal.
The company reports 54,000 API users and more than 500 billion daily tokens, but those figures are company claims and do not establish paid retention or gross margin.
The business resembles a Chinese attempt to occupy the traffic and policy layer between models and applications, where the economic value may come from aggregate workflow knowledge rather than raw inference.

Open at Leiphone Ask Claude ChatGPT

Zhejiang humanoid robots turn trade-show demos into work chains

# Analysis Leiphone

At China’s robot conference, Zhejiang Humanoid showcased linked industrial, retail and kitchen tasks, emphasizing repeatable operation, precision and handoffs over isolated stunts.

The exhibit lengthened the unit of evaluation: an order runs from purchase to delivery, a meal from preparation to cleanup and a factory task from one robot’s handoff to the next.
The company’s architecture separates world modeling, long-horizon planning and low-level vision, force and joint control, matching the different timescales of physical work.
Reported deployments include cloth handling with 2 mm precision and 90% plus system success, chemical dispensing within 1 mL and a 2,000-unit apparel-robot order; these are company claims.
The strategic shift is from asking where a robot can appear to asking how long it can stay useful without human rescue, a much harsher test of return on investment.
Chinese coverage presents trade shows as an industrial acceptance test, where success rate, cycle time, deployment duration and transfer across sites matter more than theatrical capability.

Open at Leiphone Ask Claude ChatGPT

Beijing is writing the carmaker onto the ticket, and Tesla will not be there to collect

The manufacturer now answers for the violation

Hundreds missing at the Gyirong crossing

Washington stops interviewing, Huawei calls Cairo

Alibaba shipped the Qwen4 blueprint early

Podcasts stopped asking what AI can do

Threads we are pulling