---

Today's main theme is the re-layering of the compute narrative. On one side are Nvidia's earnings test and questions about cloud vendor dependency; on the other are alternative paths brought by non-Transformer architectures and local LLM hardware, with giants starting to place their bets individually. In autonomous driving and embodied AI, Luxeed V9's L3 timeline and the PhanthyMotus co-construction plan are progressing, and edge-side AI competition is shifting from phones to desktops. Below is the full version of 24 source articles.

---

🧠 I. AI Large Models

Departed founders unveil new architecture, non-Transformer LLM debuts — Reuters reports that some founders and investors of the Prometheus project, previously backed by Bezos, have parted ways and launched a large model based on a non-Transformer architecture, clearly challenging the mainstream technical path. Since its proposal in 2017, Transformer has been the unified foundation for mainstream LLMs, with its status rarely challenged over eight years, and controversies surrounding it never ceasing—some dislike the computational overhead of attention mechanisms, others bet on more compute-efficient alternatives. Team background and product details are currently limited, with more information pending official response.

This divergence has moved from lab debates to entrepreneurial action. If alternative architectures can truly produce usable models, the industry's valuation logic will need to be reassessed. In the short term, this poses long-term narrative risks for Nvidia and cloud vendors betting on Transformer compute stacking, while being thematically positive for new compute and inference optimization companies. For investors, this adds a non-consensus exposure, maximizing both risk and imagination.

Xiaomi AI Cube prototype unveiled, multi-chip local LLM targets Apple Mac — Xiaomi revealed the AI Cube engineering prototype, using a multi-chip combination to run LLMs locally, focusing on edge inference, low latency, and data privacy. Foreign media view this as a frontal assault on Apple's Mac local AI ecosystem. Xiaomi's previous edge AI layout focused on phones and car cockpits; AI Cube expands the battlefield to desktops and workstations.

Edge AI competition is upgrading from phone assistants to local compute boxes, with terminal manufacturers starting to build workstation-like categories. Inference cost and bandwidth anxiety are pushing compute downward, providing marginal benefits for edge chip solution providers and local model supply chains. Beyond Apple, there is now another heavyweight player in high-cost-performance local compute. Multi-chip solutions offer flexibility but come with increased system integration complexity; at the prototype stage, cost advantages are not yet apparent.

• IBM open-sources Granite 4.2 reasoning model family — Officially released with technical analysis, breaking down data, architecture, and training details item by item. The Granite series is IBM's flagship model line for enterprise customers; version 4.2 focuses on reasoning capabilities, with weights and code public. Auditable reasoning chains and controllable outputs are core selling points for enterprise compliance scenarios. The open-source community can fine-tune and deploy independently, while enterprise clients can choose as needed. The arms race for enterprise-grade open-source models has shifted from closed to open source.
• Anthropic launches $5 million grant program — Soliciting proposals from independent researchers to assess AI's impact on well-being, with research conclusions expected to be made public. Assessment dimensions include usage duration, dependency, and changes in well-being, with grants disbursed in batches. Top labs are starting to put social impact assessments on the table, using real money to buy third-party perspectives rather than talking to themselves behind closed doors.
• GitHub releases pre-production LLM evaluation guide — The core reminder is direct: high benchmark scores don't guarantee success in production. The guide covers the complete process from building evaluation sets and selecting metrics to pre-launch regression testing, concluding that evaluations should focus on real scenarios and failure cases, not just leaderboard scores. For most teams, building custom evaluation sets often takes longer than tuning models; this guide breaks the process down into an executable checklist.

---

💻 II. AI Software

Zoom gives conservative guidance in earnings — Product lines continue to expand, covering meetings, customer service, and AI companions comprehensively, but growth expectations are flat, and post-market sentiment is cool. Over the past year, Zoom packaged AI meeting summaries, intelligent customer service, and collaboration tools into its product matrix. Expansion efforts are significant, but monetization conversion is still early. The conference call maintained the basic business of video conferencing, with AI product lines contributing limited revenue.

The market has limited patience for AI monetization in the video conferencing sector; no matter how good the expansion story, revenue curves must catch up first. This guidance pours lukewarm water on the AI commercialization expectations for the entire collaborative office sector.

• BOSS Zhipin Q2 net profit 1.945 billion yuan, up 173.1% YoY — Revenue 2.399 billion yuan, up 14.1% YoY. AI matching efficiency significantly improved profit margins. AI matching and automated operations reduced unit acquisition and service costs, showing profit elasticity. The AI transformation of recruitment platforms is moving from concept to numbers on financial statements.
• IBM reviews construction of production-grade multi-agent systems — Technical teams listed permission boundaries, evidence-based decision-making, and human-machine collaboration as the three major difficulties. Pitfall experience is worth more than architecture diagrams. Permission boundaries determine what data agents can access; evidence-based decision-making requires traceable records for every operation; human-machine collaboration answers which steps must remain with humans. The distance from demo to production for multi-agents is longer than most companies imagine.
• Instagram launches First Draft — Automatically edits Reels footage into highlight clips, so creators no longer need to pick shots frame by frame. The feature is rolling out gradually. Editing parameters automatically match footage rhythm, but creators can still manually adjust the final cut. AI editing officially enters mainstream social products, lowering the barrier for short video production further. Automation in content creation and distribution is happening simultaneously.
• Apple iOS 27 imminent, Siri AI likely to use waitlist mechanism — MacRumors reports that all three new systems will feature Siri AI. After the official release, initial usage scale is expected to be controlled via a first-come, first-served basis. The waitlist approach aligns with Apple's previous practice of phased rollouts. Specific rollout pace depends on performance after the official release, with the initial scope likely focused on the US region.
• Huawei MatePad Pro older tablets receive HarmonyOS 6.1.0.135 update — System package approx. 2.75GB, enhanced AI photo retouching color bleeding function, and added preset apps for Education Space. Education Space integrates learning tools for student scenarios, while enhanced AI retouching benefits daily users. Older devices continue to receive edge AI updates, as HarmonyOS's AI capabilities settle onto existing devices.

---

🤖 III. Humanoid Robots

Paradigm joins UBTECH and others to launch PhanthyMotus co-construction plan — The first general-purpose embodied Agent base moves from open source to multi-party co-construction. Over ten embodied manufacturers, including UBTECH, collectively join to jointly decide data, interface, and deployment standards. The launch event was held in Beijing, with initial member companies opening joint tests within the year. A single company cannot sustain a complete ecosystem; co-construction is becoming the new play for embodied intelligence. The battle for base standards has started early; whoever gets their ecosystem running first holds the next round of discourse power.

---

🚗 IV. Autonomous Driving

Luxeed V9 deliveries exceed 20,000 units, L3 upgrade available next year — Richard Yu announced that cumulative deliveries of Luxeed V9 have exceeded 20,000 units, setting the fastest delivery record for NEV MPVs in the 500k RMB segment. Retail market share in the high-end MPV market in July was approx. 68%. Zhao Changjiang revealed that flagship hardware is pre-installed, allowing OTA upgrades to L3-level intelligent driving next year. The critical point for switching responsibility entities is entering the product timeline.

High-end MPVs are a niche segment rarely tackled by domestic NEVs before; the 500k RMB price range has long been dominated by imported and joint-venture models. Luxeed V9 broke through with intelligent driving capabilities and delivery speed. From assisted driving to L3, Huawei's ADS first validated mass production in high-end MPVs, directly catalyzing the Huawei Car BU supply chain and partners like BAIC. Expectations for high-level intelligent driving penetration rates have been raised another notch.

• Voyah pushes major update for Huawei ADS 5 — Models like Taishan X8, FREE+, and Zhilin Huawei Edition are among the first to upgrade, covering core scenarios like urban navigation, highway navigation, and parking. Pushes are batched; owners should refer to official notifications. ADS 5 is Huawei's latest software version for intelligent driving; after the update, navigation assistance and urban scenario capabilities are enhanced simultaneously. The diffusion of Huawei's intelligent driving solutions among mainstream automakers continues, and the debate between self-developed and external supply routes persists.
• Xiaomi Auto Pengcheng N90 Max smart electric door supports door-opening collision warning — Official Q&A addressed details like anti-door-killing, rear seat-to-bed backrest load capacity, and track table storage. Electric doors monitor vehicles approaching from the side-rear before opening and stop automatically if obstacles are detected. For safety, officials recommend keeping system versions updated to get the latest collision avoidance strategies; related functions will iterate via OTA. Scenario safety for electric doors is becoming a new detail in smart vehicle competitiveness.

---

🌐 V. Physical AI

Google Research releases AgentHands — Research generating interactive gestures for XR spatial dialogue agents allows virtual proxies to make natural gestures coordinated with gaze and voice rhythm during introductions, directions, and demonstrations, no longer stiff balloon figures. Gestures are inferred directly by language models without preset motion libraries, offering better generality than traditional motion capture solutions. Anthropomorphism of spatial agents advances further, with XR interaction research shifting from recognition to expression.

• Study says AI won't replace radiologists but will significantly change their workflow — Reading processes, workload, and doctor roles are being reconstructed. AI handles initial screening, while doctors shift to review and complex cases. The study breaks the workflow into three stages: initial screening, marking, and reporting. The medical imaging segment, which has the highest degree of automation, is seeing mature human-machine division of labor first. For institutions, AI-assisted screening also alleviates the long-standing shortage of radiologists.

---

📈 VI. Macro & Market Data

Nvidia earnings test, cloud vendor dependency becomes biggest variable — This week's earnings shift market focus from quarterly performance to the depth of bundling with hyperscale cloud vendors. Cloud vendors account for the bulk of Nvidia's revenue; any wording regarding client capex pace and AI ROI will be amplified into a directional indicator for the entire sector. Nvidia is the profit engine of this AI rally; the market now demands both growth and proof of returns.

A-share mapping stocks in the GPU supply chain are also waiting for this guidance. Weakness pressures Innolight and Foxconn Industrial Internet; strength adds fuel to the compute sector. The post-earnings conference call is equally important; management's comments on order visibility will be scrutinized word by word. During earnings week, Nvidia's wording alone dictates the weather for the whole market.

• **Barc