Dalan Technology, a startup focused on long-range imaging solutions, has raised tens of millions of U.S. dollars across its angel and Pre-A rounds within a month.

The funding was led by HSG (Sequoia China) and FutureX Capital, with existing investor Jiacheng Capital continuing to participate. Proceeds will be used for product R&D, supply chain development, and team expansion.

The company was founded by Sun Zhe, former head of products at EZVIZ (Hikvision’s consumer brand). Its core team includes engineers from Huawei, DJI, and vivo.

Dalan focuses on long-distance imaging, leveraging computational imaging and system-level design to deliver professional-grade telephoto performance without relying on large optical lenses, enabling more compact devices.

At CES 2026, its flagship product won an official “Best of Media” award and attracted coverage from outlets including Forbes, TechRadar, and PCMag.

The company is targeting a large global market, including an estimated 300 million wildlife enthusiasts and 1 billion sports fans, as well as use cases such as concerts, hunting, and astrophotography.

Source:36Kr

MRTIS TechBio (Beijing) Co., Ltd., an AI-driven nanomedicine company, has received regulatory clearance from the China Securities Regulatory Commission (CSRC) to pursue an overseas listing in Hong Kong.

The company plans to issue up to 273 million shares, while 39 existing shareholders will convert approximately 852 million domestic shares into tradable H-shares under a full-circulation scheme.

Founded in 2020 by Dr. Hongmin Chen (U.S. National Academy of Engineering member) and MIT scientists Caidan Lai and Wenshou Wang, MRTIS focuses on AI-enabled nanomaterial innovation.

Its proprietary NanoForge platform supports large-scale lipid nanoparticle (LNP) design, with a database of tens of millions of compounds. The platform enables targeted delivery to eight key organs and tissues, including liver, lungs, and immune cells.

Its lead candidate, MTS-004, is the first AI-enabled formulation drug in China to complete Phase III trials. Another candidate, MTS-105, has received U.S. FDA orphan drug designation and is being developed as a potential first-in-class mRNA-encoded TCE therapy for solid tumors.

The company has raised more than $350 million (≈ RMB 2.5 billion) from investors including XtalPi, HSG, 5Y Capital, and China Life.

Source：IPOzaozhidao

TARS has completed a $455 million Pre-A funding round (approximately RMB 3.3 billion), setting a new record for the largest single-round financing in China’s embodied AI sector. Within just one year, the company has rapidly emerged as a leading player in the industry.

This marks the second time TARS has broken a sector financing record. In Q2 2025, the company secured $242 million (approximately RMB 1.75 billion) in what was then the largest angel round in China’s embodied AI history. One year later, TARS has once again raised the bar, achieving both the largest Pre-A round and the largest single-round financing ever recorded in the sector.

The round was significantly oversubscribed, underscoring strong confidence from top-tier domestic and global investors. The investor base spans four key categories:

• Financial investors: The round was co-led by Hillhouse Ventures and HSG (formerly Sequoia China), with participation from Meituan Long-Z Fund, CICC Capital, Kailian Capital, Oriental Fortune Capital, and Junshan Investment.
• Strategic investors: Meituan Strategic Investment increased its stake as a cornerstone investor and co-led the round. Existing investors including Qiming Venture Partners, Linear Capital, BlueRun Ventures, Xianghe Capital, and Hongtai Aplus also doubled down.
• Industrial investors: TCL Capital, Future Capital, Shoucheng Holdings (0697.HK), C&D Emerging Investment, Hengxu Capital, and China Auto Investment joined the round, supporting commercialization across diverse application scenarios.
• State-backed capital: Beijing Robotics Industry Development Investment Fund and Shanghai Guotou Pioneer Fund jointly invested in an embodied AI company for the first time, aligning with China’s “15th Five-Year Plan” for robotics and supporting TARS in becoming a leading player in regional embodied AI ecosystems.

Source：IPOzaozhidao

Lin Junyang, the former Alibaba Qwen technical lead and one of the youngest P10-level engineers at the company, has founded Pragmatik Labs in Shanghai with a 2 billion dollar valuation at angel stage. Gaorong and HSG co-led the round with 100 million dollars each, Tencent invested 20 million, and the Shanghai Future Industry Fund participated. The valuation was reached without any product, user base, or revenue disclosure, an early-stage pattern more common at the top of the AI talent market but still notable inside China.

Pragmatik Labs, whose Chinese name is Yuyong Technology, is publicly framed as a next-generation agent company spanning digital and physical worlds. The website describes work on reasoning, tool use, environment interaction, and long-horizon task execution, with two business lines: digital agents for knowledge work and physical agents for real-world environments. No product has been announced. Lin personally controls most of the equity through related entities, with external investors holding 12 percent of the domestic operating company.

The career arc behind the valuation is the unusual part. Lin joined Alibaba's DAMO Academy in 2019 after his master's, worked on M6 and OFA pre-training models, and became Qwen's technical lead in late 2022. Qwen launched its first open-source model in August 2023 and grew into a family covering language, code, vision, and multimodal. By January 2026, more than 200,000 Qwen-derived models existed globally with cumulative downloads above one billion. Lin resigned on March 4 and published a long-form essay on March 26 titled From Reasoning Thinking to Agentic Thinking, which set the conceptual direction for Pragmatik Labs.

The pivot from foundation models to agents is not a generic rebrand. Lin's argument is that the next leap in AI capability comes from systems that interact with environments over long horizons, call tools, take actions, and adjust based on feedback, rather than from larger static language models. While at Qwen, he had already set up a robotics and embodied intelligence team in October 2025. Both threads are now embedded into Pragmatik Labs.

The pricing reflects a broader capital pattern in which top AI founders can be valued before shipping. Investors cite Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, which raised 2 billion dollars at 12 billion dollar valuation in early stages. Pragmatik Labs will need a product to justify the current number. For Lin, the test is no longer whether a Qwen-class model can be built at all, but whether an agent company can be built from scratch by a single architect.

AOE Tech Labs, the Chinese team behind the agent desktop product Floatboat, published complete Harness evaluation data on August 7 across five third-party agent benchmarks. The base model was DeepSeek-V4-Flash, the cheapest credible model on the market. With the model and cost held constant, only the execution Harness was swapped. The same model that scored below Claude Opus 4.8 on all five benchmarks on DeepSeek's own Harness went on to beat Opus 4.8 on all five inside Floatboat's Harness.

The pricing makes the result more striking. Opus 4.8 lists at 5 dollars per million input tokens and 25 dollars per million output. At the 3:1 input-output ratio common in agent scenarios, Floatboat's blended cost is 0.175 dollar per million tokens against Opus 4.8's 10 dollars, a 57.1 times gap. The five benchmarks are third-party public leaderboards, including OpenAI-built BrowseComp. The data is not a question-selection artifact.

The methodology is the most rigorous part. The control used the official DeepSeek-V4-Flash-0731 release, not a preview checkpoint. The control group ran on the provider's own Harness in minimalist mode, which makes the comparison stricter because the provider's engineering was in the room. The Floatboat side ran in isolated physical sandbox environments with identical input parameters. The single variable was the Harness itself.

Ordered by task horizon from short to long, the gains are non-decreasing: 1.9 percent, 9.6 percent, 12.6 percent, 19.9 percent, 23.6 percent. Short tasks are essentially one-shot Q&A, where model quality dominates. Long-horizon tasks, which is what real work looks like, are dominated by the execution system: file access, tool round trips, state persistence, multi-step self-correction, and rollback. Real work lives at the long-horizon end.

AOE Tech Labs introduced the Harness Leverage Ratio, or HLR, to make the comparison auditable. The formula is the gain from changing only the Harness divided by the gain from upgrading to a stronger reference model. HLR greater than 1 means the Harness alone beats the model upgrade. On DeepSWE, Floatboat reported an HLR of 3.57: keeping the model identical and switching only the Harness produced 3.57 times the public performance span represented by upgrading to Opus 4.8. Across the five benchmarks, HLR climbed monotonically with horizon length from 0.78 times to 3.57 times.

AOE Tech Labs was founded in late 2025 with HSG and VLight Capital as seed-round investors. The team shipped Floatboat Desktop in January, FloatIM in April, and FloatSchedule in May. The thesis is that a system continuously closing the gap between an evolving execution target and the model's drifting output is the unlock for daily-use agents. The cheapest model can now match or beat the most expensive frontier model, when the Harness is built for long-horizon work.

In the early hours of August 12, Lin Junyang announced on X the founding of Pragmatik Labs, based in Shanghai, pointing toward next-generation agents. Explaining the name, he wrote that he studied linguistics because a friend recommended it, then turned to computational linguistics and NLP, and that the name means returning to where things happen, and also pragmatism: we believe AI should aim for that. A few months after leaving Alibaba's Qwen team, Lin has formally become an entrepreneur. Investors including Gaorong Ventures, HSG, and Tencent appear on his financing list, with market reports placing the valuation at 2 billion US dollars.

Born in 1993, Lin holds a bachelor's from China University of International Relations and a master's from Peking University. He joined Alibaba DAMO Academy in 2019 and moved into the core Qwen team, leading development of the Qwen series into globally influential open-source models. On March 4 he announced his departure with me stepping down, bye my beloved qwen.

Lin's startup lands at a notable moment: China's technology industry is raising the position of scientists. More researchers are entering the cores of model teams, foundation model leaders sit close to the highest decision-making structures, and young researchers leaving big companies can quickly obtain capital on technical judgment. The knowledge scientists hold is becoming scarce industrial power again. These scientists have systematic training in computer science or AI, graduate research experience, or long-term frontier research. In industry, they stand between frontier research and company decisions, judging technical direction when no ready answer exists and turning research into engineering paths. The scarce value extends to resource allocation: compute, R&D budgets, and teams follow technical judgment.

China has seen this power shift before. When technology matures, markets, channels, and scale rise; when industry re-enters uncharted territory, those who find new answers move toward center stage. Similar changes are happening across companies: Baidu placed Wu Tian, head of foundation model R&D, directly in Robin Li's reporting chain; Tencent brought in Shunyu Yao, a Tsinghua Yao Class undergraduate and Princeton PhD, to lead large models and AI infrastructure; ByteDance has Wu Yonghui, a Peking University undergraduate and HKUST PhD, leading Seed foundation model research. Among startups, Kimi founder Yang Zhilin holds a Tsinghua bachelor's and Carnegie Mellon PhD, and MiniMax founder Yan Junjie earned his PhD from the Chinese Academy of Sciences Institute of Automation.

The pattern is visible across the industry: whether at large firms or startups, the technical arc of the AI era is putting researchers back at the center of power. China's major model ecosystem increasingly resembles a scientist's era, where frontier researchers who can see where the field is going also control how resources are deployed to get there.

Recent milestones show non-invasive brain-computer interfaces expanding beyond electroencephalography. On Aug 26, Wuhan University's Renmin Hospital reported the world's first high-resolution semi-invasive retinal BCI clinical application, and days earlier Tianjin University and United Imaging released "uMR Shen Guan," a full-stack magnetic-resonance BCI solution that integrates imaging, hardware, and decoding.

The field splits into invasive, semi-invasive, and non-invasive routes. Non-invasive systems keep sensors outside the skull, exchanging some spatial precision for no surgery, reusability, and portability. Inside that track, traditionally EEG-dominated, functional ultrasound and other signal-collection methods are drawing research attention as the industry pushes toward higher spatial resolution, multimodal decoding, and closed-loop feedback.

Commercial paths are diverging. Medical rehabilitation is the clearest: defined indications, device registration, and hospital demand make it the most straightforward business. Neuracle has built EEG-VR rehabilitation systems and motion-imagery VR rehab tools, serving hospitals including Huashan, Xuanwu, and the China Rehabilitation Research Center. Its prospectus, filed with the Shanghai exchange, shows more than 20 medical-device registration certificates covering EEG machines, EMG and evoked-potential instruments, and transcranial electrical stimulation devices, plus tES-EEG and TMS-EEG combinations. In 2025 it ranked first domestically in EEG machine shipments and shipment value.

Sleep and brain-state management target a broader, high-frequency consumer base, but still need to prove long-term efficacy and retention. MindMatrices builds a portable polysomnography device, a temporal-interference deep-brain stimulation system, and a closed-loop sleep intervention around sleep rhythms and neural modulation. Human-computer interaction offers the widest space yet remains the furthest from stable revenue, since keyboards, touchscreens, voice, and eye tracking already dominate; the near-term niches are assisting people with severe paralysis and acting as an input layer in hands-occupied industrial settings, with AR-VR and embodied AI as longer-term complements.

Scaling will hinge less on technology concepts than on engineering stability, clinical evidence, regulatory clearance, user retention, and data governance. Non-invasive BCI's most plausible first big market, industry observers say, is medical rehabilitation, where the path from registration to hospital procurement is already visible.

MetaX disclosed its 2026 half-year report on Aug 31, posting revenue of 1.324 billion yuan, up 44.67% year over year, and a net profit attributable to shareholders of 612 million yuan — reversing a year-earlier loss of 186 million yuan. The figure that better reflects underlying profitability looks less rosy: deducted non-recurring net profit was still minus 49 million yuan, though the loss narrowed by 75.83%.

The three other members of China's much-watched group of four domestic AI chip vendors — Moore Threads, Biren, and Enflame — have all released first-half results, and none has crossed into full profitability yet. Moore Threads reported 1.736 billion yuan in revenue, up 147.41%; Biren reached 1.236 billion yuan, up 1997.6%; and Enflame posted 1.12 billion yuan, up 279.08%. Enflame begins its public subscription process on Sep 2, issuing 43.03 million new shares for a roughly 10% post-listing stake and targeting 6 billion yuan, completing listings for all four firms.

The four take different technical paths. Moore Threads, MetaX, and Biren follow a general-purpose GPU route aligned with the CUDA ecosystem, while Enflame has chosen a dedicated-architecture (DSA) approach with its in-house TopsRider software platform, which is not CUDA-compatible.

The commercial window is opening as demand rises. Per IDC, China shipped about 4 million AI accelerator cards in 2025; Nvidia still led at roughly 55%, but Chinese vendors supplied about 1.65 million units, a 41% share, with Huawei's HiSilicon the largest domestic supplier and Alibaba's chip division second. Nvidia cards have cooled in China under U.S. export controls, pushing more AI labs — DeepSeek, Zhipu AI's GLM, and Moonshot's Kimi among them — to adapt their models to domestic accelerators including Huawei Ascend, MetaX, and Moore Threads. Consulting firm CIC projects China's AI accelerator market to exceed one trillion yuan by 2028, with domestic solutions reaching about 90%.

Research intensity remains a drag on profitability: MetaX spent 39.65% of revenue on R&D in the half, Moore Threads 44.30%, and Biren 65%. The real test, analysts note, is whether any vendor can string together R&D, chip performance, software ecosystem, and customer orders into a consistently profitable chain.

DeepSeek has released its first native multimodal model. DeepSeek-V4-Flash-Vision-Exp went live on Hugging Face on Aug 31 under an MIT license, marking the V4 series' initial support for image input alongside text.

The release packages the model files, a tokenizer, a prompt-encoding reference implementation, and a minimal PyTorch inference implementation covering the vision encoder, an aligner, DFlash attention, a mixture-of-experts routing layer, Hyper-Connections, and DSpark modules. DeepSeek flagged the weight as an experimental "Exp" build; the model has been available on the DeepSeek API since Aug 21.

The vision capability lets the model describe pictures, read text in screenshots, and analyze charts, accepting JPEG, PNG, GIF, and WebP inputs. On pure-text tasks — agents, reasoning, and world knowledge — the company says performance is on par with the release version of V4-Flash, preserving earlier strengths. On agent benchmarks requiring visual understanding, it gains substantially over V4-Flash and brings multimodal agent capability close to Claude Opus 4.8.

By open-sourcing the inference stack as well as the weights, DeepSeek is positioning the model to slot into diverse agent frameworks and tooling, extending its open-weight strategy from text into vision-enabled work. For developers and enterprises, the release lowers the barrier to deploying a competitive vision-capable model locally or on standard infrastructure, consistent with the wider push across the industry toward cost-efficient, customizable open models.

Vision-capable models have become central to the AI agent market as developers pair screenshots and real-world imagery with reasoning. DeepSeek's move follows a pattern of increasingly open, multimodal releases from Chinese labs, which have used permissive licenses and aggressive pricing to gain adoption across U.S. developer platforms, cloud catalogs, and enterprise model menus. Because the model is lightweight for a frontier-class weight and ships with a minimal inference reference, teams can deploy it quickly without proprietary orchestration tooling.

The company has said it will keep iterating on the V4 line, with the vision-exp build serving as a validation vehicle for architecture choices — the aligner, DSpark, and Hyper-Connections components — before a broader multimodal release. For now the experimental flag signals caution: recognition and chart understanding are strong, but DeepSeek advises testing against specific workloads rather than assuming production-grade behavior across all image tasks. Early adopters, from research groups to application developers, are using the open stack to evaluate whether the model's near-Opus-level multimodal agent skills hold up in their own pipelines.]]> contact@pandaily.com (Pandaily) https://pandaily.com/moonshot-ai-kimi-k3-us-cloud-revenue-share-aug2026 https://pandaily.com/moonshot-ai-kimi-k3-us-cloud-revenue-share-aug2026 Tue, 01 Sep 2026 02:24:27 GMT

Moonshot AI has, over the past month, sat on two American tables at once. One is policy: on July 22, U.S. Treasury Secretary Scott Bessent said the department was considering adding the company to a trade blacklist, while technology official Michael Kratsios accused it of distilling Anthropic's Fable to build Kimi K3 and obtaining Nvidia GB300 servers. Moonshot denied the claims, saying K3's gains came from original architectural work.

The other table is business. According to a Reuters report on Aug 26, Moonshot is in separate talks with Microsoft, Amazon, and Google to put Kimi K3 on Azure, AWS, and Google Cloud, hoping to claim up to 30% of revenue generated by K3-related services on the three clouds. A deal would mark the first significant revenue-sharing agreement between a Chinese AI company and major U.S. cloud providers.

The negotiations reflect a broader progression. Chinese open-weight models first crossed into U.S. enterprise channels through partners such as IBM, whose watsonx.ai directories list DeepSeek-derived models and whose recent $240 million agreement with Together AI includes DeepSeek, Kimi, and MiniMax in its model menu. Kimi's second act is not just getting in, but figuring out how the money is split once inside.

Kimi K3 is a 2.8-trillion-parameter open-weight model that is expensive to run, with Moonshot recommending clusters of 64 or more accelerators, which pushes large-scale consumption toward hosted clouds. Its license already reserves a commercial boundary: any company and its affiliates running a model-as-a-service business generating more than $20 million in trailing 12-month revenue must reach a separate agreement before commercial use.

The model's capability, per Artificial Analysis cited by Reuters, is now comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8, at a reported cost around a third of Anthropic's Fable. On top of an earlier token revenue-sharing and joint innovation deal signed with ChineseSoft on July 20, a cloud agreement would give Moonshot a revenue stream that scales with overseas usage. That matters as it prepares for an IPO at a valuation reportedly up to $50 billion — and as Alibaba weighs a similar revenue-sharing model for its next open-weight Qwen.

At the World Robot Conference 2026, SynapX chose a different story from building a rival to Tesla's Optimus. The embodied startup — founded just seven months ago, with roughly 1 billion yuan raised and a unicorn valuation — introduced three product lines covering a brain, a hand, and a data-collection stack, redirecting attention from what a robot looks like to how silicon-based production gets defined.

Founder and chief executive Du Dalong, formerly the sixth employee at Horizon Robotics and an early member of Baidu's deep learning lab, argues that the two essential building blocks of embodied AI are not the body but a generally capable brain and a hand that can act causally on the physical world. A brain without a hand stays stuck on a screen; a hand without a brain only repeats industrial motions.

The products map directly onto that thesis. SYNWorld is an embodied-native world model, introduced at a billion parameters with a next version planned this year and a move toward the trillion-parameter range slated for next year. OctoH-Hand is a high-degree-of-freedom bionic dexterous hand. OctoSense combines a fisheye headband, an EMG wristband, and an exoskeleton data glove to capture first-person vision, posture, and force data from people working naturally.

The model is built on a data flywheel: human operation generates the torque, pressure, and touch signals video cannot record, and collecting it during ordinary work is how embodied training scales. SynapX does not make full robots or run terminal scenarios. It positions itself as a base layer, selling the brain, hand, and development platform to companies building embodied products, and has described the offering as a "middle-school graduate silicon workforce" that can be scenario-trained on the job.

Du Dalong places embodied AI's "ChatGPT moment" at the point a robot makes a hamburger better than a person, a task requiring force control, multi-finger coordination, and delicate contact sensing. He expects that window within one to two years, and notes Tesla spends roughly half its hardware research budget on the hand for Optimus. Once a robot makes a hamburger, he argues, food-service scenarios — sandwiches, wraps, and more — unlock at once, and physical skills become a scalable, low-cost utility.

Imvision Innovation, an AI imaging chip company spun out of SenseTime, has become one of China's most closely watched names in computational photography — and its progress raises a pointed question: can it move beyond its current position at the edge of Ambarella's territory?

The company inherited part of SenseTime's smart-life business, an AI sensor and AI-ISP line that took shape in 2021. Its V1 chip shipped to customers including Sony, Transsion, and vivo, and V2 strengthened AI performance. After being carved out as an independent, self-funded chip company in 2024, Imvision found its opening when Insta360 shipped the Ace Pro 2 in October 2024 with a dual-chip design: a 5nm AI chip handled overall performance while a dedicated imaging chip managed noise reduction and detail optimization. Insta360 then invested 30 million yuan in Imvision in 2025.

Rather than attacking Ambarella's full system-on-chip head-on, Imvision entered as a co-processor — a "second chip" for new AI workloads — then used flagship devices to validate its algorithms, packaging, software, and manufacturing. The strategy buys time, but not much room.

Ambarella, founded in 2004 by Wang Fengmin and Les Kohn, has built a two-decade platform spanning video processing, image processing, and codecs. It powered GoPro's HERO line and now pushes AI with the CV5, CV3, and the 4nm CV7, backed by research spending that exceeded $226 million in fiscal 2025.

Behind Imvision loom platform players — SigmaStar reportedly holds about 26.7% of the global vision-AI SoC market by 2024 shipments, and Horizon's spin-off is pushing endpoint vision — while top integrators like Insta360 are pulling chip-definition back in-house.

Chief executive Zhang Jiehua, who helped found DJI's chip team and worked at STMicroelectronics and Nvidia, is betting the company can fold AI, ISP, encoding, and main control into its own SoC before customers merge the work themselves. The window may span only a product generation or two.

At the HarmonyOS Ecosystem Conference 2026, held on Aug 28, the Global Intelligent IoT Consortium (GIIC) unveiled the "Hongtu Plan," an ecosystem program developed with Huawei that channels funding and process support into the development, certification, and commercialization stages of the OpenHarmony chain.

The move turns cross-device ambitions into quantifiable incentives. The OpenHarmony community now counts more than 13,000 contributors, over 650 ecosystem partners, and more than 2,000 products, while devices running HarmonyOS 6 topped 80 million by Aug 20. The plan aims to make that scale durable by easing the hardest stretch — the path from code to shipped goods.

Development incentives target OpenHarmony modules and development boards for small and standard systems. Qualifying models receive 480,000 yuan for small systems and 1.5 million yuan for standard systems, once they pass compatibility testing and reach commercial shipment readiness; each partner can claim up to three per category per year.

A second track targets certification. Unified testing, expected to go live by the end of September and coordinated by the OpenAtom Foundation, GIIC, and Huawei, lets partners complete OpenHarmony compatibility, unified interconnect, and HarmonyOS Connect tests in a single submission and earn three certificates — cutting test cycles and costs by roughly 30%.

The third track rewards shipping volume: lightweight systems earn 2 yuan per unit after cumulative shipments of 1 million, small systems 8 yuan after 200,000 units, and standard systems 30 yuan after 50,000 units, with per-partner caps of 4 million yuan per category and 8 million yuan in total.

Concentrating subsidies on foundational modules and boards — where the industry most lacks off-the-shelf options — is meant to expand the pool of reference platforms device makers can call on. Brands including Cofoe, Niu, and Skyworth showed OpenHarmony-based consumer products on site, while manufacturers such as Goldcard Smart, Hualong Xunda, and ArcherMind joined the ecosystem. Looking ahead, OpenHarmony plans to fuse the operating system with AI, positioning it as a foundation for the intelligent-connectivity industry.

A testing report from the China Academy of Information and Communications Technology (CAICT) has put a new name on the local-model map. StartLux, a Shanghai-based AI company behind the StartLux-V1.0-27B-Preview, took second place overall in the MCP special test of the trusted-AI benchmark lineup — ahead of DeepSeek-V4-Flash — using just 27 billion parameters.

The MCP test evaluates six specialized tasks — location navigation, web search, browser automation, financial analysis, code repository management, and 3D design — plus a comprehensive assessment, focusing on multi-tool coordination, complex task execution, and interaction in real environments. Models tested included DeepSeek-V4-Pro (1.6 trillion parameters), DeepSeek-V4-Flash-0731 (284 billion), Step-3.7-Flash (198 billion), StartLux-27B-260715 (27 billion), Qwen-3.6-27B (27 billion), and AgentCPM-Explore (4 billion).

StartLux-V1.0-27B-Preview scored 39.25 for second place, beating both the 284-billion-parameter DeepSeek-V4-Flash and the 198-billion Step-3.7-Flash. At the same parameter size it edged Qwen-3.6-27B by 5.34 points. It ranked first in location navigation, and tied or took first on browser automation and financial analysis — in some cases matching the trillion-parameter DeepSeek-V4-Pro.

The model is built on Qwen3.6-27B with targeted post-training enhancement. StartLux developed what it calls an "AI trains AI" (Auto Research) approach, running training experiments autonomously and refining strategy through feedback — which the company says is the first use of the method for a local agent model in China.

The result underscores a broader shift. With frontier-model performance clearing practical thresholds, the parameter-race era has matured into a homogenized phase; user priorities increasingly center on solving real problems, data security, and cost control rather than benchmark scores. Global players are moving the same direction, with Google's Gemma 4, Meta's open Muse Glimmer, and Nvidia's Nemotron 3.5 Lightning all aimed at local deployment.

StartLux-V1.0-27B-Preview runs on consumer PCs, and the company plans to launch its first generation of local intelligence solutions this year. The CAICT result sends enterprises a signal that compact, locally deployable models can now hold their own against far larger clouds on the tasks that matter in real workflows, an argument beginning to reshape procurement decisions.

China's large language models led the world in weekly inference volume for the 18th consecutive week, according to a National Business Daily analysis of OpenRouter data for the Aug 24-30 period. Global AI model traffic reached 113 trillion tokens, up 21.11% week over week.

Chinese models accounted for 55.16 trillion tokens, up 36.26%, while U.S. models totaled 17.07 trillion tokens, up 76.71%. For the first time this week, the top three spots on the ranking all belonged to Chinese models.

Leading the way was an anonymous model called Ox Alpha, nicknamed "Niu Lai" in Chinese developer circles, which climbed to No. 1 on 15.7 trillion tokens, up 36%. Launched on OpenRouter on Aug 20, it offers a context window of roughly 1.05 million tokens, supports text, image, and video input, and was free to preview.

Independent researcher Ben Davis tested Ox Alpha on 10 subtasks of the DeepSWE software engineering benchmark, where it recorded about an 80% pass rate, ahead of Claude Fable 5 (65%), GLM-5.3 and Grok 4.6 (62%), and GPT-5.6 Sol (52%).

On the evening of Aug 26, Zhipu AI publicly claimed Ox Alpha, confirming it as GLM-5.3-Flash — the first natively multimodal release in the GLM-5 family — and open-sourced the model. It is priced at one-tenth of GLM-5.3, with a limited-time discount to one-twentieth. In its launch week, the model entered the chart at No. 6 on 6.16 trillion tokens.

The previous week's leader, DeepSeek-V4-Flash, slipped to second on 12.3 trillion tokens, while Xiaomi's MiMo-V2.5 held third place for a second week at 9.14 trillion. Gemini 3.7 Flash made its debut at 3.95 trillion tokens, up 120%, priced at $0.75 per million input tokens and $3.75 per million output until Dec 31.

Falling off the chart this week were GLM 5.2 and DeepSeek-V4-Pro-0423.

The latest data cap a stretch in which cost-efficient Chinese models have steadily drawn global developer traffic, reinforcing a shift in how inference demand is distributed across the market rather than a purely competitive race between companies.

Suzhou Chipsens Technology, a Chinese MEMS device maker, has released a standard "public-edition" micromirror large-array chip, a step toward localizing the core components behind optical circuit switching (OCS) and supporting self-reliant AI compute infrastructure.

OCS has gained traction as AI compute demand pushes computing clusters toward supernode architectures and faster interconnect. Optical circuit switches route data as light, avoiding the electrical-to-optical conversions of electronic switching and offering high bandwidth with low latency. Google uses the technology heavily, equipping its seventh-generation TPU Ironwood clusters with more than 2,000 OCS switches each, while NVIDIA's Spectrum-XGS aims to build an Ethernet ecosystem around it.

Several technical routes compete, including MEMS, digital liquid-crystal, piezoelectric and optical-waveguide designs. MEMS is the most mature and holds more than 70 percent of the OCS market because it balances port expansion with cost, and both Google and Lumentum build on it.

Chipsens, founded in 2019, controls MEMS chip design, wafer fabrication, packaging and module integration, and has focused on pressure sensing and micromirror modules for automotive, industrial, medical and consumer uses. It tape-outed a large-array micromirror for a 320 by 320 channel OCS in 2025, and in August 2026 brought the standard chip to market.

The chip integrates 416 independently rotating mirror units on silicon in a package of 33.6 by 20.85 millimeters, roughly 0.66 millimeters thick. It offers a deflection angle of plus or minus 6.5 degrees on the X axis and 5.5 degrees on the Y axis, infrared reflectivity of at least 96 percent, a response time under 10 milliseconds, a cycle life exceeding one billion cycles, better than 90 percent mass-production yield, and support for 400 by 400 port products.

By lowering the entry barrier, the standard chip lets domestic and overseas customers build tailored OCS solutions quickly for AI data centers and core networks. Switching in the optical domain can cut data-center energy use by about 40 percent, making it a candidate for reconfigurable optical layers in 100,000-card training clusters. Chipsens says it plans to scale toward 1024 by 1024 ports. In August it closed a funding round led by SMIC's investment arm, SMIC Capital, to expand micromirror capacity and advance next-generation products.

Chinese commercial remote sensing is moving from seeing the planet to listening to it. On August 30, Chang Guang Satellite Technology held a departure ceremony for the Diting 01 A, B and C satellites at its Jilin aerospace information park, with the trio set to be shipped to Jiuquan for launch. The satellites make up the first group of the Diting constellation, a planned formation that senses radio-frequency emissions across a wide spectrum and locates radio transmitters.

Instead of selling images, Diting aims to answer different questions: who is transmitting, on which frequencies, and from where. Three satellites in formation observe the same signal from separate positions, using the differences to triangulate its source. It is essentially a ground radio direction-finding station moved into orbit, and it works even when a ship switches off its automatic identification system or clouds hide its hull, as long as its radar or communications equipment remains on.

Investors have called the operator, Tianjin-based Xingkan Jiuzhou, a Chinese analog of HawkEye 360. That U.S. company launched its first three Pathfinder satellites in 2018 and has built commercial radio-frequency geolocation into a data business serving radio management, maritime surveillance and defense agencies. RF remote sensing sits closer to signals intelligence and electronic warfare than conventional optical imagery, which is why HawkEye 360 moved quickly into U.S. national-security customers.

The shift points to a broader change in China's space industry. For years the competition was about cheaper rockets, launching more satellites and building bigger constellations. Now that many more satellites are in orbit, a harder question has emerged: what data will they sell? Optical imagery is mature, synthetic-aperture radar is expanding fast, and weather, infrared and hyperspectral bands are crowding. Radio-frequency data is one of the scarcer new layers, complementing rather than replacing traditional remote sensing.

For now, Xingkan Jiuzhou's disclosed applications cover radio management, maritime and shipping, public security and railways, with no military missions announced. The significance of the Diting 01 launch is less that China gained three more commercial satellites and more that a new kind of satellite data product has entered in-orbit validation. The next round of competition, the industry argues, will favor whoever can capture data others cannot and turn it into a service customers keep paying for.

China's "four dragons" of domestic GPUs — Moore Threads, Biren Technology, Enflame and MetaX — have all now reached the public markets or are about to, and they are entering a differentiation phase. The defining question, analysts say, has shifted from whether a company has a chip to whether it can close a commercial loop.

The companies are growing quickly while narrowing losses. Moore Threads, the earliest to list, leads revenue and has moved into steady batch shipments across a general market. Biren and Enflame focus on large AI-compute cluster tenders, which offer high impact but volatile demand, while MetaX is climbing more steadily. All four channel heavy sums into software, the field analysts see as the real battleground for taking share from NVIDIA's entrenched CUDA ecosystem.

The four split into two technical camps. Moore Threads, Biren and MetaX follow the general-purpose GPU path, staying highly compatible with CUDA to lower developer switching costs; Moore Threads' own MUSA architecture has drawn more than 800,000 developers. Enflame is the outlier, choosing a dedicated architecture (DSA) that does not mimic CUDA, backed by its proprietary Tops Rider full stack covering drivers, compilers and operator libraries.

Customer strategy also diverges. Enflame is deeply tied to Tencent, which the company has called both a moat and a cap given that Tencent bought a large share of its revenue. Biren's customer base spans national compute centers, the three state telecom operators and leading AI labs. Moore Threads sells broadly across the internet, telecom and government sectors.

Despite the buzz, the four remain a minority of the Chinese AI accelerator market. IDC counted roughly four million AI accelerator cards shipped in China in 2025, with NVIDIA taking about 55 percent and Huawei supplying nearly half of the domestic share. Consulting firm CIC expects China's AI accelerator market to exceed one trillion yuan by 2028 with domestic solutions reaching around 90 percent. In the next two years, the four are expected to settle into natural tiers across training, inference and rendering rather than face a sudden shakeout.

Chinese memory chip pioneer Yangtze Memory Technologies (YMTC) is moving toward a listing that would create China's first publicly traded 3D NAND integrated device manufacturer. The review status of its holding company's STAR Market IPO in Shanghai changed to "accepted" this week, marking a milestone in a ten-year attempt to break a global oligopoly.

YMTC was founded in Wuhan in 2016 under the National Storage Base, backed by a roughly 160 billion yuan project that the state's Big Fund called its largest single investment that year. At the time, Korean and Japanese 3D NAND had reached 48 layers and U.S. makers 32, while mainland China had not mass-produced any NAND flash. YMTC began at 32 layers.

The company then climbed rapidly: a first 32-layer tape-out in 2017, mass production in 2018, 64 layers in 2019, a 128-layer quad-level-cell device in 2020 that skipped the 96-layer step, and a 232-layer product in 2022 that placed it among the first globally to ship more than 200 layers. By 2025 it led into 294 layers, and by March 2026 three-hundred-plus-layer devices were shipping across the industry, with rivals already pushing toward 400 layers.

Technology, however, did not immediately translate into market share. YMTC held only about 4.8 percent of global NAND in 2022, when Samsung, Kioxia, Micron, SK Hynix and Western Digital still dominated. Its breakthrough came through capacity. Between 2023 and 2025 the company spent 21.3 billion, 31 billion and 36.2 billion yuan respectively on plants and equipment even as the memory industry endured a downturn, pairing expansion with price cuts to seize share while incumbents cut back.

By the second quarter of 2026, Counterpoint Research estimated YMTC at 14 percent of global NAND capacity share, behind only Samsung's 25 percent and SK Hynix's 22 percent, and TrendForce data put it third worldwide and first in China on sales and shipments in early 2026. The company still warns its scale trails the giants and needs sustained investment. Yet it has moved from follower to co-runner, and its next decade has just begun.

At an 80-year-old man's birthday party in Japan, two Chinese humanoid robots danced on stage, then one extended an arm to hand the guest a gift. The image has become an emblem of a widening trend: China's robot makers are actively moving into one of the world's strongest robotics economies.

Unitree is now among the best-known Chinese humanoid brands in Japan, said one engineer at a Japanese humanoid software firm, with customers increasingly finding the company rather than the other way around. According to Smart Analytics Global, Chinese manufacturers accounted for 97 percent of global humanoid shipments of about 18,500 units in the first half of 2026. Aging Japan, where demand is a focus across warehousing, restaurants, homes, carmaking and airports, is a key destination.

Japan's industrial robotics base remains deep. Its makers dominated the traditional "big four" and supplied roughly 45 percent of the world's industrial robots in 2020 by exports, yet newer service and logistics scenes demand equipment that is easier to deploy and adapt. McKinsey projects Japan's available labor will shrink by about 1.5 million people over the next two decades, and a Reuters-Nikkei survey found 34 percent of 220 firms already using, planning, or considering AI robots.

Chinese companies, however, are finding the market is not an easy shortcut. Japanese clients reward automation that plugs into existing processes, is maintained by local teams, and shows a clear return — not just a striking humanoid story. Warehousing provider Geek+ placed its first overseas project in Japan in 2017; Pudu's catlike BellaBot delivery robot entered restaurant chains, and home-automation brand SwitchBot now serves more than two million Japanese households. Hitachi partnered with UBTECH in May 2026, and Unitree signed a Japan agency deal with GMO AIR in June.

Distribution runs through trading houses and systems integrators, and sales cycles are slow, often taking years. Keenon set up a Japanese subsidiary in 2022 with more than 200 support points, while Dobot opened a Tokyo office in 2023. Japan's robot industry booked 1.0456 trillion yen in 2025 orders, up 25.7 percent. For Chinese players, Japan remains a test of patience and long-term commitment.