Guangzhou-based robotics firm Huaruizhipu unveiled Praxis One, a physical-AI edge compute platform built on MediaTek's Genio Pro 5100 system-on-chip, at an industry event alongside the 9th China Robot Summit in Hangzhou on September 17. The product is framed as an embodied edge compute brick for heavy-duty humanoids and high-compute robot bodies—not a smartphone story—pairing on-device multimodal and vision-language-action workloads with real-time motion control on a single chip.

Huaruizhipu contrasts the design with common split stacks that put a cognitive accelerator beside a separate motion microcontroller. Praxis One integrates high-level decision-making and low-latency control on one SoC to cut cross-chip latency, data loss, and board complexity. A lighter sibling, Praxis Nano on MediaTek Genio 720, targets mid-size robots and motion-plus-sensor closed loops, while Praxis One is aimed at multimodal perception, VLA policies, lightweight world models, local large-model inference, and complex task planning on heavier platforms.

MediaTek and partners cite Genio Pro 5100 as a 3-nanometer Arm v9.2 all-big-core octa-CPU design with an eighth-generation NPU rated at about 106 TOPS, plus support for up to 16 virtual FHD camera channels, Wi-Fi 7, Bluetooth 6.0, and optional 5G. On-device Qwen3-7B inference is quoted around 23 tokens per second, intended to keep scene understanding and dialogue available when cloud links drop. Industrial wide-temperature support from about −40°C to 105°C is pitched for factory and outdoor deployment.

Huaruizhipu also outlined adjacent pieces of its stack—edge data-collection hardware branded under its RuiMou line, a physics-constrained latent world model, and scene robots—plus turnkey packages for OEMs and labs. Official English branding for the company remains lightly documented outside Chinese materials and the harmony-robot.com site; this report uses Huaruizhipu for the corporate name and Praxis One for the product mark pending a clearer global trademark confirmation.

Researchers at the Institute of Semiconductors of the Chinese Academy of Sciences and several collaborating research institutions have built a sliding ferroelectric tunnel junction that delivers a giant tunneling electroresistance ratio while keeping ultrahigh endurance, with related results published in Science. The work targets nonvolatile memory and compute-in-memory arrays where readout contrast and cycling reliability usually trade off against each other in compact two-terminal cells.

Two-dimensional sliding ferroelectrics such as 3R-MoS2 and 1T′-ReS2 switch polarization by interlayer slip rather than ion displacement, avoiding defect buildup that limits conventional oxide ferroelectrics and supporting fatigue life beyond 10^11 cycles. Their remanent polarization, however, is two to three orders of magnitude weaker than oxide counterparts, so two-terminal devices often cannot produce a strong electrical readout without aggressive fields that shorten device lifetime.

The team used van der Waals heterostructure interface engineering to stack Cr / h-BN / 3R-MoS2 / monolayer graphene—later extended to 1T′-ReS2. Hexagonal boron nitride supplies a uniform, low-leakage tunnel barrier; the sliding ferroelectric sets a durable built-in potential that modulates barrier height; graphene's low density of states and weak screening let polarization also gate the carrier density involved in tunneling. That dual lever converts a weak remanent polarization into a large resistance contrast without sacrificing the fatigue advantage of sliding switching.

At a 0.5 V read bias, the device reached a tunneling electroresistance ratio of 1.9×10^7—more than four orders of magnitude above prior sliding ferroelectric junctions—while delivering 222 A/cm² on-state current density and endurance beyond 10^11 switches. Reliable switching held down to 13-nanosecond pulses at 6.5 fJ/bit. An 8×8 array showed consistent bistable nonvolatile behavior across all 64 cells, underscoring integration potential for dense storage and crossbar in-memory compute designs.

Alibaba DAMO Academy, working with the First Affiliated Hospital of Zhejiang University School of Medicine and other clinical partners, has open-sourced DAMO RADAR, a general-purpose abdominal contrast-enhanced CT model whose results appear in Science. The system is designed to flag more than 146 abdominal findings across 18 organs in a single pass, with reported accuracy reaching expert radiologist level on the evaluated set—an explicit break from one-disease, one-model medical imaging AI that struggles in messy real clinics.

DAMO said conventional vision-language learning struggles on sparse CT volumes, so the team used organ-level fine-grained alignment: three-dimensional scans are decomposed into anatomical units so images and report text match at the organ scale, then adaptive contrastive modeling adjusts the training signal. The approach aims for scalable, multipurpose, and more interpretable diagnosis without requiring extra manual labels for every new condition as the finding list grows.

In nearly 40,000 real-world exams, DAMO RADAR reached an AUC of 0.913 across the 146 assessed findings. In a reader study against 26 radiologists from multiple hospitals, its average accuracy exceeded that of 23 physicians. With AI prompts, doctors' sensitivity rose about 10% and reading time fell more than 30%, and junior readers approached senior-level performance, according to the published results. On out-of-distribution acute abdomen cases outside the initial training emphasis, AUC remained about 0.904.

Code and weights are available on GitHub under the Alibaba DAMO Academy organization, with the Science paper detailing methods and trials. Hospital partners describe the model as a navigational aid for one of radiology's hardest workloads—multi-organ abdominal CT—while DAMO frames RADAR as a step from specialty detectors toward general medical imaging foundation models that can transfer beyond the abdomen CT setting as clinical validation expands.

SenseTime has released SenseNova U1.5, an 8-billion-parameter Mixture-of-Transformers model that treats understanding, reasoning, and pixel-space generation as one native multimodal system. Unlike pipelines that perceive with a vision encoder and synthesize through a separate variational autoencoder, U1.5 maps pixels and text into a shared backbone without those external modules, following the company's NEO-unify line while tightening spatial reconstruction for higher fidelity at production resolutions.

The architectural shift replaces independent patch MLP decoding with a lightweight spatial decoder: visual tokens are reshaped into a two-dimensional feature field, then restored through Pixel Shuffle stages and local convolutions so neighboring regions exchange information before pixels are finalized. SenseTime said the interface still maps each 32-by-32 region to one visual token, yet supports native generation up to 4K with fewer seam and texture discontinuities than the prior U1 design, aided by resolution-aware noise conditioning across wider aspect ratios.

Post-training follows a specialize-then-unify recipe. Separate reinforcement-learning experts target visual aesthetics, bilingual text rendering, infographic layout, and image editing, then multi-expert on-policy distillation folds those skills into a single student along its own generation trajectories. SenseTime reports stronger results on open image-generation and editing suites versus U1, including competitive bilingual text rendering and multi-reference editing, while keeping multimodal understanding scores broadly intact on standard STEM, VQA, and OCR benchmarks.

Alongside the technical report on arXiv and Hugging Face Papers, SenseTime is open-sourcing training code covering supervised fine-tuning, reinforcement learning, and on-policy distillation, with model assets pointed to SenseNova and Hugging Face collections under the OpenSenseNova organization. The release is the full U1.5 product stack—not the earlier U1.5-Lite-Preview—aimed at developers who want a compact native unified vision model they can inspect, fine-tune, and retrain end to end.

Huawei Cloud will commercially launch its Ascend 950 Lingqu AI cluster cloud service in China on September 30 and open it to global markets on November 30, Huawei Cloud CEO Zhou Yuefeng said at Huawei Connect 2026. The dates mark a cloud SKU timeline for Ascend 950 compute rather than another SuperPoD architecture brief: customers can book a UnifiedBus-linked 1,024-card cluster as a managed cloud offering on a published calendar.

Huawei said the Ascend 950 system uses its UnifiedBus interconnect—marketed under the Lingqu name—to assemble a 1,024-accelerator pod delivering up to 1 EFLOPS of FP8 compute and 2 EFLOPS of FP4, with 256 TB of globally addressable unified memory. The design is intended to support end-to-end training of large models without extra cluster-level partitioning or adaptation, keeping a thousand-card job coherent under one memory address space.

The cloud commercial window sits alongside broader Ascend roadmap updates shared at the same event. Rotating Chairman Wang Tao said more than 1,000 Ascend supernodes are already deployed and that Ascend 950 supernodes have entered scaled commercial use. Ascend 960DT is planned for the first quarter of 2027 and Ascend 960PR for the third quarter, with a stated cadence of roughly one major Ascend generation per year thereafter, including longer-range 970 and 980 milestones.

For buyers, the news value is the concrete China-then-global calendar plus the published cluster envelope: 1,024 cards, EFLOPS-class FP8/FP4 throughput, and 256 TB unified memory under UnifiedBus. Huawei Cloud is framing Ascend 950 Lingqu as a ready-to-consume AI cluster service rather than a paper architecture, with domestic availability at the end of September and international availability two months later for teams that want supernode-scale training without assembling the fabric themselves.

CATL has restated an engineering milestone for all-solid-state batteries: small-batch production is targeted for 2027. In a September reply on the Shenzhen Stock Exchange interactive platform, the company said it continues firm investment in the all-solid-state route and that technology sits at an industry-leading level, with small-batch output expected that year. The statement is a manufacturing roadmap marker, not a claim of mass-market vehicle penetration.

That timeline sits beside a more cautious technology-readiness framing from chairman Robin Zeng at the June 2026 Summer Davos forum. Using a 1-to-9 Technology Readiness Level scale, Zeng placed today's all-solid-state cells around TRL-4—laboratory principle validation—and said million-unit vehicle installation before 2030 remains unlikely. Industry reporting links the 2027 goal to manufacturing readiness near levels 7–8 (process capability for limited runs), including roughly 2 GWh of pilot capacity planned at CATL's Yibin base with equipment commissioning targeted in the second half of 2026.

Engineering bottlenecks explain the gap between small-batch and scale. Sulfide solid electrolytes are moisture-sensitive and can generate hydrogen sulfide, so lines need ultra-dry environments that reuse little of today's wet-process liquid-electrolyte tooling. Solid-solid electrode interfaces, warm isostatic pressing yields, and dry-electrode coating remain open process risks. Cost structures are still dominated by sulfide electrolyte precursors such as lithium sulfide, which keeps all-solid-state cells several times more expensive than mainstream liquid lithium-iron-phosphate packs.

Peer roadmaps converge on a similar cadence: semi-solid or hybrid cells for 2026 vehicle demos, all-solid-state small batches around 2027, and broader commercial volumes only after further yield and cost work into the next decade. For CATL, 2027 is best read as the start of controlled pilot output and A/B-sample vehicle validation—an event-driven climb from TRL-4 toward manufacturable cells—rather than the moment all-solid-state becomes the default traction battery.

OPPO used its 2026 Developer Conference in Zhuhai on September 17 to unveil ColorOS 17 as an on-device AI operating-system stack rather than a skin refresh. The company said ColorOS now exceeds 770 million monthly active users across more than 90 markets, and that OPPO, OnePlus, and realme will take the first synchronized upgrade wave. ColorOS 17 is slated to ship first on upcoming Find X10-series and OnePlus 16 devices, with the same continuum spanning phones, tablets, wearables, cars, and smart-home hardware.

The compute pillar is On-Device Compute, an end-side large model built on a first-party Linear Attention architecture. OPPO claims native 128K context with about 48% lower memory use and 55% lower energy versus the prior generation—aimed at privacy-preserving, low-latency assistants that stay on the handset. Persona X, described as a memory co-evolution engine, tracks data, environment, and behavior signals so the system can record, forget, and reflect over long companionship rather than reset each session.

Agent Matrix upgrades the Breeno assistant brain with longer context and stronger intent parsing; OPPO cited a 19% rise in Breeno beta satisfaction under device-cloud collaboration. Everyday hooks include one-sentence access to about 200 Alipay services, Tencent Map trip planning, and WeChat messaging or calls. Breeno Suggestions partners with more than 40 third-party services across 700-plus micro-scenarios, with a claimed 95% prediction accuracy from OPPO's swarm sensing stack.

Under the UI, Aurora Engine unifies rendering to cut render load by about 30% and peak launch memory by up to 25%, while Tide Engine personalizes keep-alive so full app retention rises 55.6% versus the previous release. Fluid Design ties lock screen, Dynamic Island-style surfaces, and home transitions into continuous motion without trading smoothness for spectacle. The product story is proactive OS intelligence—Linear Attention on device, Persona X memory, and an agent matrix—not another handset SKU launch.

Alibaba's Qwen team has launched Qwen3.8-Omni-Flash, a native omnimodal model that jointly handles text, image, audio, and video with a 1-million-token context window. The model is live on the Qwen AI platform as of September 18, positioned less as a captioning demo and more as an agent stack for long audio-video workflows—planning tasks, calling tools, and delivering finished media assets.

On a suite of about 30 public and internal benchmarks, Qwen3.8-Omni-Flash scored more than 26% higher on average than the prior Qwen3.5-Omni-Plus generation. Gains were especially large on agentic audio-video and long-horizon tasks: WildClawBench-MM rose 36.5 points, AgenticVBench rose 22.3 points, and UniClawBench reached 69.6. Core perception also improved—LongAudioSpan by 8.3 points and OmniVideoBench by 9.6—while AliMeeting diarization error rate / concatenated word error rate fell from 88.11 / 89.61 to 3.35 / 17.18.

Agentic long-video understanding is a practical focus. Rather than scanning every frame, the model can decide what to watch and listen to, concentrating tokens on relevant spans. On OmniVideoBench, that agentic mode lifted accuracy from 63.4 to 67.8 while cutting token use from about 145,736 to 79,117—roughly 45.7% fewer tokens. Meeting workflows accept up to one hour of audio-video input for speaker separation, transcription, minutes, and follow-up actions such as email or coding via tools.

To support those pipelines, Alibaba expanded Qwen-MM-Plugins and open-sourced Qwen-Live Harness for real-time continuous multimodal interaction. A Realtime SKU adds low-latency streaming response, accent-aware oral practice, and spatial audio cues that estimate sound direction and distance. API pricing for hourly audio input was cut by more than 98%, and hourly audio-video input by more than 93%, according to the company—framing the release as both a capability jump and a lower-cost path for production omni agents.

Lexiang Technology, also listed in English corporate materials as JoyIn Technology, put its Aether embodied model through a public outdoor stress test on September 16 in Shanghai's Minhang district. Two humanoid robots ran a barbecue-stall service loop for more than an hour—taking orders, coordinating kitchen work, plating, and responding to live customer changes—while cameras livestreamed the session as a continuous open-environment trial rather than a clipped demo.

Aether is positioned as a roughly 4-billion-parameter cross-embodiment model. Lexiang says training used about 200 hours of human video and zero hours of real-robot demonstration data, relying on an energy-based state evaluation stack, streaming memory for long tasks, and an explicit self-model of the robot body so the same policy can drive different morphologies. An earlier uncut dual-robot room-cleanup clip under the MVP codename had already drawn industry attention to collaborative recovery when direct actions failed.

In the barbecue livestream, one robot handled front-of-house confirmation and guest service while the other managed grill-side preparation; mid-order changes such as reducing spice required the greeter to re-brief the cook and keep water and napkins moving. Lexiang framed the hour-plus outdoor run as a harder public verification than studio demos because stalls, lighting, and spectator interference cannot be fully scripted, and failures accumulate on camera in real time.

The company has not yet published a full technical report that would let third parties reproduce the zero-robot-data claim or the advertised zero-shot success rates, and a single livestream cannot settle broader debates about self-evolution marketing. What the September 16 session does establish is a concrete public artifact: a dual-robot, cross-embodiment outdoor service task lasting over an hour under live scrutiny—useful evidence to weigh alongside lab demos until independent benchmarks arrive.

Xiaomi's MiMo team has opened a public livestream dashboard for reinforcement-learning post-training of MiMo-V2.6, exposing parallel Pro and Flash runs at mimo.xiaomi.com/rl/. Instead of shipping only weights or a blog post, the page streams step counts, reward curves, token throughput, accumulated cost, and selected infrastructure events—an unusual degree of operational transparency for a frontier-scale agent RL experiment.

According to the published setup, each training step processes on the order of 2 billion tokens across 1,568 prompts with 16 asynchronous rollouts per prompt. The run mixes multi-task agent environments and harnesses in a single job and emphasizes heavier grader compute, including agentic in-group credit assignment with test-case and rubric-style rewards. The pipeline is fully asynchronous so rollout generation, environment execution, reward scoring, and parameter updates need not share a rigid lockstep.

Early public tallies cited across coverage put combined Pro and Flash spend above about $1 million within roughly a day and a half after Pro started on September 15, with Pro averaging on the order of $20,000 per hour and Flash around $10,000 per hour at the snapshots reported—figures that illustrate burn rate, not a fixed future tariff. Mid-run DeepSWE and reward metrics have been posted while training continues; they should be read as interim signals, not final rankings.

Team lead Luo Fuli said related technical details would be open-sourced in coming weeks. The livestream is distinct from Xiaomi's earlier Robotics-U0 embodied world-model release: this is language/agent post-training observability, not a robotics checkpoint drop. Outside labs will still question whether dashboard numbers fully prove live continuity and how much any hidden distillation traffic matters, but the public RL console itself is the news—cost, steps, and rewards shown in the open while MiMo-V2.6 is still learning.

Zhipu AI has opened GLM-5.3-FlashX on its API and experience center, billing the tier as a high-speed inference cut of the Flash family with peak generation claimed near 200 tokens per second. The company says the speedup rests on inference capacity from roughly 100,000 domestic AI accelerators plus further investment in infrastructure and serving optimization—positioning FlashX as a production latency product rather than a new base-model launch.

The base checkpoint, GLM-5.3-Flash, was open-sourced on August 26 as a 320-billion-parameter mixture-of-experts model with about 18 billion active parameters and a 1-million-token context window. It previously appeared anonymously as Ox Alpha on public routing platforms. Z.ai's engineering post describes building a from-scratch production serving stack for Flash on that domestic-accelerator cluster, citing limited per-chip memory and bandwidth, incomplete kernels, and multimodal long-context traffic as the hard constraints.

Much of the adaptation work, Zhipu says, was carried out by an Infra Agent powered by GLM-5.3: engineers set objectives and review critical changes while the agent proposes diagnoses, kernel patches, and stack edits under dense feedback from correctness tests, traces, and end-to-end metrics. Reported outcomes include roughly 3× end-to-end throughput versus the initial same-hardware baseline and per-token efficiency approaching mainstream NVIDIA GPU serving economics—framed as an RSI-adjacent production loop in which the model helps optimize the system that serves the model.

FlashX therefore lands as the public speed SKU on top of that domestic-accelerator story. Independent operators will still want to reproduce the 200-token-per-second ceiling and cost claims on their own traffic mixes, but the combination of an open MoE Flash base, a large domestic serving cluster, and an agent-assisted inference stack is the concrete product narrative Zhipu is putting forward for teams that need lower latency on the same Flash family.

ChangXin Memory Technologies (CXMT) is preparing a NAND flash research-and-development production line at its new Beijing site, according to people familiar with the plans cited by Reuters and follow-on reports on September 18. The DRAM specialist would also lean on a Beijing research institute whose remit includes NAND development—an expansion that would push CXMT beyond its core DRAM business into flash memory at a moment of global shortage.

Industry executives expect AI-server demand to keep memory tight through at least 2027, with SK hynix's chief executive previously warning that next year could be especially constrained from a supply perspective and TrendForce projecting NAND tightness to ease only in the second half of next year. Against that backdrop, a Beijing NAND R&D and pilot line would give CXMT a path to broaden its customer set without yet committing publicly to full-scale commercial flash fabs.

Inside China, CXMT and Yangtze Memory Technologies (YMTC) are often described as the dual stars of domestic memory: CXMT leads in DRAM while YMTC anchors NAND. A CXMT flash push would put the two on overlapping turf and add a domestic contender alongside global peers such as Samsung Electronics, SK hynix, and Micron. Sources stressed that timing for the R&D line's start and whether pilot work graduates to mass manufacturing remain undecided.

The angle is category strategy and capacity learning—not an IPO or fundraising hook. A separate Beijing DRAM fab expansion discussed earlier this summer should not be confused with this NAND R&D plan. Pandaily's September 17 note on CXMT HBM trial yields and YMTC XStacking hybrid-bond DRAM stacks covered stacking process maturity; this story is about CXMT entering the NAND product category itself. Until CXMT discloses process node, wafer starts, and commercial volumes, the Beijing flash line remains a reported R&D capacity bet amid a multi-year shortage cycle.

On September 10, Alibaba's Amap (AutoNavi) released ABot-Earth 0.7, which it calls the first 3D-native urban world model. Built on Amap's spatiotemporal data and a 3D-native architecture, the system aims to generate continuous, enterable, and interactive digital-twin worlds from planetary to street-view scale, with coverage claimed across more than 196 countries and regions. CEO Guo Ning framed Amap spatial intelligence as understanding the world through topology, 3D structure, flow, and time—not language alone.

Unlike digital-earth products that paste satellite imagery, aerial photos, or point clouds onto a globe, ABot-Earth 0.7 is trained to form a native 3D understanding and emit city scenes end-to-end as 3D Gaussian Splatting (3DGS). Amap says a single satellite image or text prompt can produce a kilometer-scale 3D city on one consumer GPU in about 10 minutes—roughly a 1,000× efficiency gain versus traditional reconstruction pipelines—while the model fills continuous space so users are not locked to pre-collected camera paths.

Scale continuity is part of the product pitch: one model is meant to move from planet to city fabric to landmark street detail while keeping road networks, blocks, and building groups coherent, then refining facade geometry and materials at close range to photo-like quality. An experience site is live, and capabilities appear in Flying Street View 2.0, where users can joystick-roam complex buildings and large scenic areas in 3D.

Amap situates the release inside a three-layer spatial-intelligence stack—3D expression, dynamic sensing of traffic and commerce, and spatiotemporal reasoning for routing and prediction—with ABot-Earth as the entry that turns Earth from a browse-only globe into a generative, interactive world model. Soft-window coverage around September 10 presents vendor claims; independent teams will still need to measure generation quality, geometric fidelity, and the 10-minute consumer-GPU workflow outside Amap's demos.

TaichuAI has open-sourced ZDTaichu5.0-9B, an approximately 9-billion-parameter multimodal foundation model aimed at general visual understanding, spatial reasoning, agentic tool use, and embodied-AI research. The stack pairs a Qwen3.5-9B language backbone with a C-RADIOv4-H vision encoder, accepts text, single or multiple images, and video at any resolution, and supports context lengths up to 128K tokens. TMTPost summarized the September 15 release as focused on physical-world spatial perception, cross-view transforms, and embodied task planning.

On the public model card, TaichuAI reports first-tier standing among compared ~10B general VLMs for broad vision while extending into spatial, embodied, and agent settings. Highlighted spatial and embodied scores include ViewSpatial 62.50, MMSI-Bench 47.20, MindCube-tiny 78.27, ERQA 48.00, and RoboSpatial 56.00. On agent and instruction suites under the card's evaluation notes, TAU2-Bench reaches 87.70 and Claw-Eval general average 71.40, with IFEval at 93.70. Entropy-gated adaptive recurrent reasoning is described as allocating extra latent refinement steps to harder tokens.

Capability coverage spans OCR and document understanding, visual math, fine-grained 2D relations, multi-view association, 3D scene and perspective taking, multi-image and video tracking within the long context window, and multi-step tool-use planning—while tool execution remains the host application's responsibility. Spatial training topics on the card include relative relations, dense counting and boxes, camera motion and depth ordering, egocentric versus allocentric views, and high-level affordance and action planning for VLA-style adaptation.

Weights are on Hugging Face as TaichuAI/ZDTaichu5.0-9B under the NVIDIA Open Model License, retaining Qwen3.5 Apache-2.0 notices. Serving uses a custom vLLM 0.26.0 branch and Docker image from TaichuAI for OpenAI-compatible endpoints, with recommended sampling presets for spatial grounding versus general tasks. As with other vendor-led model cards, outside labs should treat benchmark leads as reported under the stated prompts and judges until third-party replications land—but the combination of a mid-size open multimodal checkpoint, spatial/agent emphasis, and ready vLLM packaging gives researchers a concrete artifact to evaluate.

On September 16, ByteDance's Volcengine said Doubao-Seed-2.1-pro updated to the 0915 release is fully available on Volcano Ark APIs, with Doubao Work upgraded in parallel and TRAE already connected. The company frames the refresh around production multimodal coding, more reliable agent delivery, and lower image and video inference token cost—reported as more than 30% below the prior generation—rather than a funding or consumer-chat story.

Multimodal coding is the concrete product hook. ByteDance argues that enterprise intent often lives in screen recordings, UI sketches, and operator habits rather than clean documents. In an official case, Doubao-Seed-2.1-pro 0915 read an undocumented Java ERP of about 280,000 lines from a recording and a few sketches, inferred architecture, and generated a working mobile front end—shifting the interface from humans writing machine-readable specs toward models reading human visual expression.

A second showcase used the open-source Luanti game repository—about 387,000 lines of code and 1,000 historical issues—with official patches held out. The model orchestrated multiple sub-agents for nearly 36 hours of root-cause analysis and cross-file repair; about 83% of tasks met a mergeable standard meant to reflect engineer-acceptable quality. Separately, the model rebuilt an interactive seasonal courtyard 3D scene from four design images, staging structure, water, vegetation, and camera effects.

Volcengine also stresses agent reliability and cost. Prior benchmarks placed Doubao-Seed-2.1-pro in a leading cohort on Terminal Bench 2.1, SWE-Pro, SciCode, OSWorld, MobileWorld, and MMMU-Pro; the 0915 cut adds stronger evidence tracing, source authority checks, and verification for long research reports, plus denser document and engineering-drawing parsing. Token efficiency improvements for image and video inference are positioned for high-frequency enterprise API use. The release is a model-and-product update on Volcengine—not a financing announcement—and outside teams will still want to reproduce the ERP and Luanti merge rates on their own stacks.

Also at Huawei Connect 2026, rotating chairman David Wang introduced OceanStor M900 AI memory storage for hyperscale inference SuperPoDs. The product targets the memory wall that appears when context windows stretch into the million-token range and multi-turn agent workloads inflate KV Cache beyond HBM and DRAM economics. Huawei positions M900 as PB-class shared memory space that lets AI infrastructure move from compute-centric design toward compute–network–storage co-design.

On Lingqu, OceanStor M900 pools and tiers KV Cache so SuperPoD memory can spill from device memory to SSD. Huawei says a single cluster can reach 64 PB of capacity, lifting per-NPU usable KV Cache from the gigabyte class into the terabyte class and raising cache hit rates for long-context reuse. That framing treats storage as an extension of inference memory rather than a conventional backup or object tier.

Performance claims center on a CPU, network, and disk-controller three-in-one architecture with native KV semantics. Huawei says NPUs can reach SSD in one hop without protocol translation or CPU forwarding, cutting access latency from the millisecond range to about 60 microseconds—roughly a 90% reduction—and delivering about 40 TB/s of aggregate cluster bandwidth, described as about 1.5× prior industry approaches. In a typical AI coding inference scenario, Huawei claims token throughput can double while time-to-first-token halves.

Cost control leans on KV-Aware adaptive scheduling that predicts KV Cache data value and lifecycle across media. Huawei claims up to 24 DWPD and about a 16× extension of SSD media life, supporting three-year stable operation with fewer drive replacements. Independent operators will still need to validate the latency, bandwidth, and endurance figures on their own SuperPoDs, but M900 is clearly positioned as agent and long-context memory infrastructure—complementary to, and distinct from, Ascend chip or NPO interconnect stories.

At Huawei Connect 2026 in Shanghai on September 17, Huawei ICT BG CEO Yang Chaobin framed Lingqu UnifiedBus as the interconnect core of a new cluster and SuperPoD co-architecture aimed at Agentic AI workloads. Huawei argued that as clusters grow, utilization often falls because cards wait on communication, while 10-trillion-parameter training and frequent agent–model exchanges push KV Cache and intermediate data far beyond single-card memory—so fabric design becomes the performance bottleneck.

Lingqu's stated design pillars start with protocol unification: more than ten interconnect protocols collapse into a single Lingqu memory-semantic fabric, with interconnect bandwidth described as moving from the 100 GB class into the TB class and round-trip latency compressed from about 7 microseconds to about 2 microseconds, enabling global memory addressing inside a SuperPoD. CPU, NPU, memory, and SSD attach as peers for decentralized access, with flexible CPU–NPU ratios and hardware-accelerated Attention/FFN decoupling for AF-separated deployment. Tiered storage pools activations and can treat DDR as a second memory for NPUs, while all-optical networking is positioned as the high-bandwidth, low-latency data highway for elastic scale-up and scale-out.

Huawei also announced layered Lingqu interconnect gear spanning cabinet, cross-cabinet, and cluster tiers. In-cabinet blades eliminate copper cable and circuit loss—Huawei claims a 4,096-card SuperPoD can save about 196 kilometers of copper. Cross-cabinet switches offer 176 ports at 1.6T each for 280T all-optical bandwidth per box at roughly 2-microsecond RTT. Xinghe UBG Lingqu network switches advertise a 1024 radix fan-out intended to support million-card SuperCluster builds as models grow toward tens of trillions of parameters.

On that fabric, Huawei described an Agentic SuperPoD cluster combining Kunpeng 950, Ascend 960 SuperPoD, OceanStor M900 memory storage, and UBG switches for heterogeneous compute and pooled resources. The same interconnect story extends to appliances: the Atlas 650E air-cooled server can dual-node Lingqu-direct-connect 16 NPUs without a switch, behaving as a small SuperPoD for on-prem trillion-parameter inference. The angle here is interconnect and system co-design—not a chip-only Ascend 960 SuperPoD near-package optics brief—and remains a vendor roadmap claim until independent cluster measurements appear.

China's industrial-robot market is no longer a four-way choice among Fanuc, ABB, Yaskawa, and Kuka alone. Industry tallies compiled by MIR DATABANK and echoed in Estun materials put domestic brands at roughly 55% of China unit share by 2025, up from under 30% around 2020, while the traditional "four families" still hold a large global installed base and retain strength in premium automotive and semiconductor cells.

On the shipment board, Estun reported about 33,400 industrial-robot units in 2025—roughly 10% share—and claimed the first all-brand volume lead in China. Inovance followed as a top domestic peer at about 9.2% overall share, with roughly 25% of the SCARA segment, building on its lead in general-purpose servo drives. Both firms are pushing vertical stacks: Estun from motion controllers and servos through robot bodies into workstations; Inovance from servo and control into SCARA, six-axis arms, and machine vision. Component localization underpins that shift—GGII-cited figures show domestic RV and harmonic reducers and servo share rising sharply through 2024.

Capability gaps remain concrete. Light-load six-axis, welding, palletizing, and collaborative robots are where domestic makers already report majority local shares and static repeatability near ±0.02–0.03 mm on some mid-tier arms. Full-vehicle body shops and advanced semiconductor tools still lean on long-certified Fanuc, ABB, and Kuka process packages. Offline programming suites such as ROBOGUIDE, RobotStudio, and KUKA.Sim embed decades of application libraries; swapping brands means migrating programs, process know-how, and engineer habits—not only the arm on the floor.

The structural read is therefore product and software depth, not a financing scoreboard. Domestic suppliers are winning new lines in lithium batteries, photovoltaics, 3C, and cost-sensitive cells, while high-end auto and semiconductor process software remains the slower frontier. Estun's early body-shop spot-welding orders mark a crack in that wall; closing it will hinge on process packages and simulation stacks as much as on axis count or sticker price.

Stable AI, working with Tsinghua University professor Peng Cui's group, has released LimiX-2, a 400-million-parameter foundation model for structured and tabular data. Weights and inference code are on Hugging Face under stable-ai/LimiX-2, with a technical report on arXiv as 2609.17488. One checkpoint handles classification, regression, and missing-value imputation in a single forward pass without task-specific fine-tuning.

Architecturally, LimiX-2 adopts Contextual Mechanism Networks (CMNs) pretrained with Context-Conditional Masked Modeling. Instead of centering on a conventional tabular prior-fitted objective that predicts a designated target given context, CMNs learn a context-dependent joint structure over features and labels—closer to discovering how variables co-generate than to fitting one column at a time. Pretraining draws on synthetic datasets from structural causal models spanning varied graphs, mechanisms, and observation processes, then scales along previously published LimiX scaling laws up to the 400M class.

On public leaderboards the team reports Elo 1935 on TabArena, 1432 on BCCO, and 1506 on TALENT—each claimed first among compared tabular foundation models and AutoGluon-style baselines under the published protocols. TabArena results also highlight strong regression and classification splits; the model card further notes causal-skeleton recovery via feature attention that encodes direct causal links. Earlier LimiX releases already covered hundreds of enterprise structured-data scenarios; LimiX-2 is the scaled CMN generation meant to push that generalist tabular lane further.

For data teams, the practical pitch is a single open weight that can classify, regress, and impute across heterogeneous tables without per-dataset retuning—useful where enterprise tables never look like public text corpora. Benchmarks should still be read as author-reported Elo under specific configs; production users will want holdout checks on their own schemas and leakage controls before replacing dedicated AutoML pipelines.

Light Origins has open-sourced LightNav-0, a compact generalist navigation model built on a Qwen3-VL-4B backbone and released with weights, code, a technical report, and the INSIGHT-Bench evaluation kit. The company frames the model as a navigation brain that shares one token interface across instruction following, open-vocabulary object navigation, and visual tracking—without task-specific prediction heads.

The data story is Real2Sim2Real at stated scale. A data engine converts more than 2,000 real-world scenes gathered from the internet into reusable simulators, then synthesizes more than 4,000 hours of aligned vision–language–action experience. Training runs in three stages: embodied-reasoning mid-training to build spatial priors, embodied supervised fine-tuning on aligned trajectories, and online reinforcement learning so the policy improves from its own failures. Camera height, field of view, and pitch are randomized so the same scenes cover multiple embodiment viewpoints before any real-robot fine-tuning.

On evaluation, Light Origins reports that monocular first-person RGB LightNav-0 leads ten navigation settings spanning VLN-CE, Matterport3D/HM3D object navigation, HM3D-OVON, and EVT-Bench tracking—ranking first among single-camera methods and remaining competitive even when panoramic systems are included. Dual-channel pointing tokens express spatial intent in image space; a three-level residual vector-quantized action tokenizer then decodes a 10-step waypoint chunk with centimeter-scale reconstruction error in the team's ablations. Zero-shot demos move the same checkpoint onto humanoid, quadruped, wheeled, and aerial platforms in unseen indoor and outdoor scenes, including crowded street tracking.

Artifacts are public on GitHub under lightorigins/LightNav-0 and on Hugging Face as LightOriginsHQ/LightNav-0 under Apache 2.0. For robotics teams, the useful claim is not another VLN score alone but a reproducible alignment loop—real scenes into sim, then back to heterogeneous bodies—aimed at reducing teleoperation-heavy navigation data collection while keeping a single RGB-and-language interface.