TRAE, ByteDance's AI development product, has launched a personal assistant for developers, which it calls TRAE Bao. The feature went live on October 29. TRAE announced it through its official channels and on its official Chinese community forum. Instead of handling one task per chat session, the assistant is designed to keep track of a developer's projects, tasks and decisions over time.

TRAE frames the problem this way: much of the time developers spend with AI coding tools goes not into writing code but into breaking down projects, assigning tasks and switching between sessions. Every new task means explaining the project background, tech stack and earlier decisions again. Complex requirements must be split up by hand, tracked across sessions and stitched back together.

The assistant uses a two-layer design. A global assistant handles memory and coordination across projects. It can read a developer's past tasks, conversations and outputs and pull out reusable solutions and constraints. When a developer starts something new and states the goal, the global assistant looks for relevant experience in earlier projects and loads the context automatically.

A project assistant focuses on one project at a time. Working from that project's context, it breaks a complex requirement into several executable tasks, runs subtasks in parallel, integrates them and summarizes the results. In TRAE's description, the developer's role shifts toward hearing reports and making decisions.

In the TRAE desktop app, the global assistant is reached through a dedicated entry point, while creating or opening a project leads to the project-level assistant. The global assistant can be linked to messaging apps such as Feishu and WeChat, so developers can send it tasks and receive progress updates without opening the client.

There are limits. The current version supports local tasks only. TRAE says cloud tasks and mobile support will arrive in later updates. Users on the official forum report that the feature is rolling out gradually, that it has appeared in the China version first, and that it currently uses automatic model selection without a manual model choice.

The launch adds TRAE to a growing group of desktop agents that let users hand off work from a phone or chat app and keep long-term memory across sessions.

SEE ALSO: Doubao Work and Tencent WorkBuddy Both Push Desktop Agents Further Onto the Phone

Tencent used its 2026 Global Digital Ecosystem Summit on October 29 to launch Buddy One, a system that ties its productivity apps for individual users to a single AI agent. Alongside it, Tencent released a new cloud storage product, Tencent Drive, designed to act as the storage layer for that agent setup.

Buddy One links Tencent Docs, Tencent Meeting, Tencent Drive and the knowledge tool ima, all running on the WorkBuddy agent. Users get one account across the apps and one shared pool of AI credits. Tencent says memory sharing between the apps will follow, and more products will join over time.

In practice, users link their accounts from WorkBuddy's PC client, version 5.9.0 or later. After that, they no longer need to tell the agent which app to search or paste context in by hand. A request triggers a cross-app search over documents, meeting records and knowledge bases the user has authorized, and the answer labels which app each source came from. Credits bought or earned in WorkBuddy become Buddy One credits that work across WorkBuddy, Tencent Docs, ima and Tencent Drive. A Buddy One VIP plan costs 700 yuan per month and bundles WorkBuddy's top tier, Tencent Meeting Pro, Tencent Docs premium membership, shared credits and 1TB of storage.

Tencent Drive is the more developer-relevant piece. Tencent describes it as cloud storage built for agents. Besides serving as the shared store for WorkBuddy, Tencent Docs and other Tencent apps, it can connect to supported third-party agents and applications through a command-line interface and the Model Context Protocol. That lets a user finish a draft with one AI tool and pick it up in another without re-uploading files. The WorkBuddy knowledge library feature already runs on it.

For businesses, Tencent Drive runs on Tencent Cloud Object Storage, with end-to-end encryption, off-site disaster recovery backup and VPC deployment. Admins can set fine-grained permissions for both employees and AI tools. Tencent also raised storage in WorkBuddy's enterprise edition from 5GB to 50GB per seat.

Tencent also said WorkBuddy will open its architecture further, letting companies plug in their own agents, third-party agents and third-party messaging systems. Tencent's cloud chief Dowson Tong said the choice of model, agent and messaging app should rest with the enterprise.

SEE ALSO: Doubao Work and Tencent WorkBuddy Both Push Desktop Agents Further Onto the Phone

DiffuSpace, the team behind the Dream 7B diffusion language model, has released and open-sourced Clover-35B-A3B-Preview. It is a diffusion language model aimed at the kind of work usually handled by conventional large language models: writing code, calling tools, browsing web pages and generating front-end interfaces.

Most language models are autoregressive. They predict one token at a time, each new token depending on the ones before it. Diffusion language models work differently. They update many positions at once and refine the output over several rounds, recovering masked or noisy text into a coherent answer. The approach promises faster generation, but diffusion models have historically been unreliable at agent tasks, with malformed tool calls and repeated failed steps.

Clover uses block diffusion. Output is split into blocks that are generated in sequence, while positions inside each block are denoised in parallel. The team says it restructured a Gated DeltaNet backbone for this block-level parallel denoising and added learnable block-size calibration. The model is a mixture-of-experts design with 35 billion total parameters and about 3 billion active per token, and it supports a thinking mode.

According to figures published by the team, Clover scores 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. On tool use, it reports 65.3 on the tau-cubed overall benchmark and 67.9 on BFCL-v4. The team says these results are ahead of autoregressive models of the same size, including Qwen3.6-35B-A3B. All of these numbers are self-reported and have not been independently verified.

Speed is the main selling point. DiffuSpace reports 515 tokens per second on agent tasks on a single NVIDIA H200, compared with 365 tokens per second for a DFlash2 speculative-decoding setup, and 680 tokens per second overall on a single GB300. In a web-agent test across 49 tasks, the team says Clover's median execution time was 1.52 seconds.

The weights are available on Hugging Face, GitHub and ModelScope under the Apache 2.0 license. DiffuSpace is also offering a hosted trial through its website and says an agent harness and a desktop coding client are in development.

Clover is labeled a preview, and its real-world reliability on long agent workflows will need outside testing. Still, it is one of the largest open diffusion language models aimed squarely at coding and agent use.

The first computing cluster built on Huawei's Ascend 950DT 1,024-card SuperPoD has been switched on in Wuhan. It was lit on October 28 at a Huawei artificial intelligence conference for Hubei province, held as part of Huawei's 2026 China roadshow.

The cluster was built with investment from Hubei Science and Technology Investment Group, a provincial state-backed investor. According to the group, the project took only one month from kickoff to power-on, a short timeline for a system of this size.

The deployment is a single liquid-cooled SuperPoD. One node carries 1,024 Ascend NPU accelerator cards. Both the compute cabinets and the interconnect cabinets are fully liquid-cooled. The system delivers 1 EFLOPS of FP8 compute and has 256TB of globally unified memory. Card-to-card bandwidth is at the terabyte-per-second level, and end-to-end communication latency is in the microseconds.

The design relies on UnifiedBus, Huawei's interconnect architecture for SuperPoDs. It lets every card in the cluster access the others directly, so the 1,024 NPUs can behave much like one large accelerator rather than many separate servers stitched together over a conventional network. Huawei positions the setup for training models with trillions of parameters and for running AI agent inference.

That direct-access design is the main point of the SuperPoD approach. Large model training spends a growing share of its time moving data between chips. Keeping memory unified across the pod and the interconnect inside the pod reduces how much traffic has to go through slower links.

The cluster will serve local demand. It is meant to give model developers and companies in Hubei a nearby source of compute for training large models and for building, tuning and running industry AI agents. Local reports describe it as filling a gap in high-end computing supply in the province.

The Wuhan launch is also an early real-world deployment of the Ascend 950 generation. Huawei first showed a physical Ascend 950 SuperPoD earlier this year and said at the time that more than 750 sets of its previous-generation Ascend 384 SuperPoD had already been deployed commercially. The 950DT cluster in Wuhan shows the newer SuperPoD design moving from showcase hardware into regional AI computing centers.

Huawei has not published the cluster's power draw or detailed utilization figures, and the performance numbers above are those given at the launch event.

SEE ALSO: Huawei Atlas 960E SuperPoD: First NPO/Hi-ONE Optical Node, 4096 NPUs, 8 EFLOPS FP8 · DeepSeek Open-Sources Ascend Versions of TileLang, DeepGEMM and DeepEP as Huawei Details SuperPoD Flex

Beijing-based commercial spacecraft company AZSpace says the structural test model of its DEAR-21 spacecraft has completed mechanical testing. DEAR-21 is the first flight model of the company's C2000 series of medium cargo spacecraft and is planned to launch and fly in the first half of 2027.

Mechanical testing checks whether a spacecraft's structure can survive the vibration, shock and loads of launch. The milestone follows the completion of the structural test model's assembly, which AZSpace announced on October 20. The schedule has moved back from earlier plans: when AZSpace showed a full-scale model of the C2000 demonstration version at a trade fair in Chongqing in May, it said the maiden flight was set for the end of 2026.

The C2000 is a step up in size for the company. Each C2000 spacecraft is designed to carry up to 2,000 kilograms, and several can be combined into what AZSpace calls version 2.0 of a commercial space station. AZSpace has said the C2000 is meant for low-cost delivery of supplies to orbit, reusable orbital transfer, and modular assembly of commercial space station sections. It is designed to stay in orbit for long periods and to support rendezvous and docking.

Founded in 2019, AZSpace focuses on commercial spacecraft and commercial space stations, with stated business lines in space station development, in-space manufacturing and space tourism. Its spacecraft under development include the B300 series of small cargo craft, which carry up to 300 kilograms each and can be combined into an uncrewed commercial space station, version 1.0; the C2000 medium cargo series; and the D6 suborbital crewed spacecraft, designed for six to nine people.

The company's first product, the A10 series, made its first flight in 2023. According to AZSpace, the B300 series completed a full commercial cycle in 2024, and in 2025 the company carried out what it describes as China's first in-orbit operation of a small commercial cargo spacecraft.

Moving from 300-kilogram to 2,000-kilogram cargo craft is the next hard step on AZSpace's path toward a commercial space station. The 2027 flight will be the first real test of whether the C2000 can do the job.

China released its first batch of extreme-weather artificial intelligence models built on Fengyun meteorological satellite data on October 29, at the 2026 Fengyun Satellite User Conference, according to CCTV News. The release is meant to move Fengyun satellite applications from describing the weather as it is toward predicting what comes next.

The first model interprets typhoon cloud imagery. It combines multichannel Fengyun cloud images with related satellite products to automatically analyze a typhoon's position, structure, intensity and likely evolution, then writes up its findings as interpretation text. The model now covers seven categories of typhoon features and draws on more than 20 real-time satellite products. It runs an interpretation every 15 minutes and needs only 27 seconds to generate the text.

The developers used this year's 26th typhoon, Surigae, as an example. After automatically identifying the storm, the model produced a combined analysis of its intensity, wind radius and rainfall characteristics to support monitoring and forecasting. Work of this kind has traditionally depended on forecasters reading satellite images by hand, so automating the first pass can give them a faster, more consistent starting point.

The second model is aimed at earlier warnings of severe convective weather such as thunderstorms. It uses the past two hours of continuous observations from the Fengyun-4 satellite and applies AI to forecast how the clouds will develop. Its key idea is to split cloud evolution into two parts: the overall movement of large weather systems, and the local growth and decay of small ones. The approach extends warning lead time to four hours.

The conference also introduced an artificial intelligence satellite meteorology technology and application innovation team at the Xu Jianmin Meteorological Satellite Innovation Center. The team brings together 49 researchers from 12 institutions and will work jointly on multimodal representation of satellite weather data and on monitoring and early warning of extreme weather.

The satellite models add to a growing set of AI weather tools in China, which so far have centered on medium-range forecasting models run on reanalysis data. These new models work directly from satellite observations, the data forecasters rely on most when a storm is already forming.

SEE ALSO: China Meteorological Administration Puts Fengqing AI Ensemble Forecast Into Operation · China's MAZU AI Weather System to Reach 30 Countries

Alibaba's workplace platform DingTalk has open-sourced DWS, the DingTalk Workspace command-line tool, and released a version that runs on Huawei's HarmonyOS PCs. The project, published on GitHub under the Apache 2.0 license, describes itself as DingTalk Workspace on the command line, built for both people and AI agents.

According to IT Home, DWS spans 25 business modules and more than 1,000 commands. The services listed in the documentation include contacts, chat and group management, calendar events and meeting rooms, to-dos, approvals, attendance records, work reports, AI tables, documents, whiteboards, cloud drive, AI meeting minutes and mail. The idea is that a user describes what they need and an agent carries it out through these commands, within the permissions it has been granted, instead of the user clicking through messages and to-dos one by one.

Agent use is a core design goal. Commands return structured JSON, and the tool ships with built-in agent skills that can be installed per product or as a single combined skill. The documentation recommends a confirmation-skipping flag for agents, a dry-run option to preview a request without executing it, and a filter option to pull out only the fields needed so agents spend fewer tokens.

Because DWS touches company data, it currently requires authorization from an enterprise administrator. The project lists OAuth device-flow login, encrypted token storage tied to the device, a domain allowlist that by default only sends credentials to DingTalk domains, enforced HTTPS, and protections meant to stop agents from being tricked by malicious input, such as path traversal and Unicode spoofing filters. DingTalk says every call passes through its open platform's authentication and audit chain.

The HarmonyOS PC support comes through an OpenHarmony build. The repository produces an openharmony/arm64 binary with an OpenHarmony Go toolchain, and every beta and stable release now builds, self-signs, verifies and publishes a static package alongside the six other platform archives. The project says the binary has been validated on a HarmonyOS PC. Some features are unavailable on that build, including SafeChat and runtime native payloads.

IT Home describes it as the first open-source office command-line tool on HarmonyOS PCs, filling a gap in productivity software on Huawei's desktop platform.

SEE ALSO: Huawei Opens DevEco Studio Public Beta on HarmonyOS PCs · Doubao Work and Tencent WorkBuddy Push Desktop Agents Onto the Phone

MoleculeMind has published a new technical report on MMDesign, its AI platform for designing proteins from scratch. The report covers 19 design projects across 18 targets and two kinds of molecules, nanobodies (VHH) and macrocyclic peptides. In 18 of the 19, the platform produced at least one molecule that lab experiments confirmed binds the target, a project success rate of 94.7%.

MMDesign runs a fixed generate, filter and validate workflow. It first generates tens of thousands of candidates computationally, then scores them on structural reliability, how natural the sequences look and a statistical-physics-based assessment. That narrows each project to between 9 and 50 candidates for wet-lab testing, with a median of 21. Of all 453 candidates tested, 139 were confirmed binders, a hit rate of 30.7%. Fifteen projects produced at least two binders, giving researchers options to compare and optimize.

The hardest test was GPCRs, one of the most important drug target families and a notoriously difficult one for antibodies, because the receptors sit in the cell membrane, change shape and expose little surface outside the cell. MMDesign obtained confirmed binders for all six GPCR targets it tried: CCR7, CCR8, EP2, GPR68, NK3R and A2AR. Hit rates ranged from 12.1% to 100%, and the best binder for every target reached nanomolar affinity, with GPR68's best at 12.1 nanomolar. All 30 tested candidates for NK3R bound, as did 10 of 11 for GPR68. On five GPCR targets with no catalogued therapeutic antibody precedent, 53 of 89 designs bound.

The report also tests repeatability. A second design round on six targets that already had binders improved hit rate, best affinity or both in five cases. For PDGFR, the hit rate rose from 4.0% to 21.1% and the best affinity improved from 150 to 37 nanomolar. For macrocyclic peptides, 16 of 21 MDM2 candidates and 11 of 28 PD-L1 candidates bound. Some candidates also showed good expression, little aggregation and low nonspecific binding in early developability tests.

Behind the designs is MoleculeMind's all-atom structure prediction model, MMFold. On the FoldBench antibody-antigen benchmark across 172 interfaces, MMFold's top-1 high-quality prediction success rate was 39.0%, about twice the 19.8% of the AlphaFold 3 comparison. The full report is available on Zenodo.

SEE ALSO: MoleculeMind Achieves Breakthrough in AI-Powered Nanobody Design with MMDesign

Apex Intelligence, a Chinese AI startup founded in June 2026, released its first model, Apex-35B-V1.0, on October 28. The weights are on Hugging Face. The model is built on Alibaba's Qwen3.6-35B-A3B and was further developed through mid-training and post-training aimed at coding, agent workflows and general tasks, as well as research work in AI, computational science and robotics simulation.

The company's pitch is less about the model's size than how it was made. Apex Intelligence says its automated research system, Apex Research, took part in generating training data, running experiments, designing the training plan, and optimizing GPU kernels and the inference system. In one example the company gives, Apex Research found that the model was weak at computational fluid dynamics tasks, analyzed where its reasoning fell short, then generated and verified new training data that was fed into later training. According to the company's internal testing, that loop raised the model's CFD-Bench score by 32.1 points. Apex Intelligence also says AI-driven optimization raised training throughput by 45% to 60% and made decoding 1.6 to 1.8 times faster.

All benchmark results published with the release come from the company's own evaluation. On that basis, Apex Intelligence says Apex-35B-V1.0 beats both Qwen3.6-35B-A3B and Recursive SuperIntelligence's Flint on all six research benchmarks it reports, outperforms GLM-5.3 on FrontierScience and PaperBench, exceeds DeepSeek-V4.1-Flash-1017 on the human-difficult split of BioMysteryBench, and posts the highest BFCL V4 score among the models it compared. On two robotics tests, RoboRSI and Real2Sim, it reports 44.5 and 21.55, close to the much larger GLM-5.3-Flash at 52.4 and 21.4.

Real2Sim asks a model to rebuild a working physics simulation from a single photograph: identify the objects, choose simulation assets with similar visual and physical properties, and recreate their layout so robot actions can be tested. RoboRSI measures how well a model controls simulated robots by writing code and revising it based on visual feedback.

The company was founded by Chen Yongchao, an assistant professor at Tsinghua University's College of AI. He frames the work as recursive self-improvement, the idea of AI systems helping to build their successors. He told Tencent Technology that AI has made clear progress in carrying out research but still lacks research taste, the judgment of which problems are worth pursuing.

SEE ALSO: DeepSeek Makes V4.1-Flash-1017 Official With MIT-Licensed Weights

Developers have spotted three model identifiers that had not appeared before in DeepSeek's official Harness repository on GitHub: deepseek-v5, deepseek-v5-open and deepseek-v5-next. DeepSeek has not announced a V5 model, and the entries are placeholders in test code, not models anyone can call.

Pandaily checked the repository. The identifiers sit in a test file for the DeepSeek account provider, packages/llm/llm-deepseek-account/tests/provider.spec.ts, at lines 154 to 156, next to the existing deepseek-v4-pro entry. The repository history shows they were added in a commit dated September 24 that moved the DeepSeek account model catalog onto the official model listing. The test checks how the client filters models by version requirement: the v5-open entry carries a minimum client version of 0.0.1, while v5-next requires 99.0.0-alpha.1, and the test expects only the flash model and v5-open to show up in the list.

Read that way, the code shows that Harness can switch on a new model from the server side, either for every client or only for clients that upgrade, without shipping a new release. It does not show that a V5 model exists or when one might arrive, and some developers believe the names are simply test placeholders. As of October 28, DeepSeek had published no announcement, model card or API identifier for a V5 series. Its current lineup is V4-Pro and V4.1-Flash, the latter released in September with a new architecture while keeping the V4 name.

The same week brought a fresh benchmark result for that current model. Artificial Analysis released Intelligence Index v5, its largest rework since v4 arrived in January, with 60% of the weighting now based on private test sets or hidden answers. Coding carries 25% of the index, and 15 points of that go to Cognition's FrontierCode 2.0, which scores code changes from 132 real open-source tasks on whether the repository's maintainers would merge the pull request.

On FrontierCode 2.0, DeepSeek V4.1-Flash-1017 at max effort scored 28%, the highest of any model with released weights. Kimi K3.1 Pro scored 25.5%, but its full weights are not yet out. Gemini 4 Argon reached 25.7% and GPT-6 Luna 25.2%.

The overall index is less flattering. DeepSeek V4.1-Flash scores 32, tied with GLM-5.3, against 58 for the leader, Claude Opus 5.5. Artificial Analysis says the gap between the best released open-weight model and the top of the table widened from 3 points under v4.3 to 12 under v5.

SEE ALSO: DeepSeek Makes V4.1-Flash-1017 Official With MIT-Licensed Weights · DeepSeek Harness v0.2.1-rc1 Arrives

Moore Threads has open-sourced Megatron-MUSA-Titan, the large-model training infrastructure it built for its own GPUs. The project targets trillion-parameter mixture-of-experts models on clusters of more than 10,000 Moore Threads GPUs and covers both pretraining and reinforcement learning. The code is on GitHub.

The company says the stack has been in continuous use for trillion-parameter MoE pretraining on its 10,000-GPU-class clusters, where long training runs achieved high model FLOPs utilization and stable training. Moore Threads has not published a specific MFU figure. The repository includes a pretraining recipe for a Kimi-K2-style 1T MoE model as a configuration example.

The central design choice is staying close to upstream. Megatron-MUSA-Titan is built around NVIDIA's Megatron-LM framework, and through torch_musa's PyTorch integration and the CUDA C compatibility of MUSA C, developers with a compatible MUSA environment can clone the official Megatron-LM, slime and miles repositories and run training on Moore Threads GPUs without modifying the upstream code. The project says general algorithms and bug fixes should go to the upstream projects, while it keeps two kinds of changes: large-scale optimizations tailored to Moore Threads hardware and clusters, and workarounds for MUSA-specific issues.

The pretraining features include blockwise FP8 training, which relies on Moore Threads' MT-TransformerEngine. MoE token dispatch and combine are accelerated with DeepEP running over the company's Async Copy Engine, which also lets expert-parallel communication overlap with computation. Other options include activation offloading with virtual pipeline parallelism and one-forward-one-backward scheduling, and pipeline-parallel communication overlapped with computation using Mooncake RDMA. A distributed trace analysis tool reads PyTorch traces from every GPU to locate bottlenecks such as pipeline bubbles and straggler nodes, and the project says it can analyze a 1T MoE run across 10,000 GPUs within hours.

For reinforcement learning, the stack supports the slime and miles frameworks, with Megatron handling training and a MUSA build of SGLang generating rollouts, so the same base serves both pretraining and RL. Standalone inference optimization is outside the project's scope.

For developers weighing domestic accelerators, the practical question is how much existing training code survives the move. Moore Threads' answer here is that the Megatron ecosystem code does not need to change.

SEE ALSO: BAAI's FlagOS 2.2 Expands Its Cross-Chip Operator Libraries · Moore Threads MTT S5000 Embodied RL Curves Match Mainstream GPUs

Iluvatar CoreX used its 2026 media open day in Shanghai on October 28 to show the Tiangai 300 general-purpose GPU running on real hardware for the first time. Until now, the company's public material on the chip had focused on architecture and roadmap slides rather than working systems.

The headline number is a performance comparison. Iluvatar CoreX says the Tiangai 300 is the first domestic general-purpose GPU whose measured performance reaches the level of NVIDIA's H200, and that third-party testing, after conversion, puts the card's overall single-card throughput at 1.05 times the H200. That figure is the company's claim. Iluvatar CoreX has not published the test method, the models that were run, or the workload conditions behind it, so the result cannot yet be checked independently.

The chip is designed to be deployed in IXL144, a supernode the company showed earlier at the China International Industry Fair. One IXL144 rack combines 144 GPUs with 18 CPUs and about 10.1TB of GPU memory, which Iluvatar CoreX describes as the highest-density single-rack GPU supernode in the industry. By its own estimate, the setup cuts the cost per million effective tokens by 40% compared with mainstream international solutions.

The open day also brought news on heterogeneous deployment. Iluvatar CoreX says it has run what it calls the first domestic AFD disaggregation scheme across GPUs and XPUs, cutting inference latency by 60%. AFD disaggregation generally refers to splitting stages of model computation, such as attention and feed-forward layers, across different types of processors so that each runs the work it handles best.

On the edge side, the company said its Tongyang TY1200 chip outperforms NVIDIA's Thor T5000 in its tests, again without publishing details. The TY1200, previously announced with 300 TOPS of dense compute for AI PCs and embodied intelligence, is also said to be the first domestic edge product to run GPU-accelerated physics simulation based on MuJoCo-Warp, a capability that matters for training robots in simulation.

Together, the demos move Iluvatar CoreX's next-generation lineup from paper specifications toward running systems. Independent benchmarks with disclosed workloads will be the next thing to watch.

SEE ALSO: Iluvatar Unveils Its Fourth-Generation Chip Roadmap

Chinese 3D generation company Hi3D has open-sourced Hi3D v1.0, a model that refines rough 3D shapes into high-resolution geometry. The inference code is on GitHub and the model files are on Hugging Face, alongside a free online demo.

Hi3D v1.0 does not create an object from nothing. It takes a coarse mesh and a reference image of the same object and refines the mesh into a detailed one. It is built on Sparc3D, a high-resolution 3D modeling framework presented at NeurIPS 2025 that focuses on representing and reconstructing complex geometry so that thin parts and surface detail survive. Hi3D keeps that core architecture and expanded the training data to about two million curated 3D assets. The team says the version was validated in its own production workflow before being opened up.

The pipeline samples sparse latents, decodes them with a self-pruning sparse VAE and extracts a mesh, then repeats. A 512 request runs one pass, 1024 runs two, and the top setting runs all three passes to reach an effective sparse extraction grid of 1536 cubed. The project notes that these figures describe nominal sparse grids rather than a dense latent tensor or a fixed triangle count. Examples in the technical report include an armored vehicle whose angled armor, tire treads and underside stay consistent from front, back and bottom views, and a winged sculpture whose individual feathers remain distinct.

The hardware bar is high. The project requires Linux, Python 3.10 and an NVIDIA GPU with compute capability 8.0 or newer, and there is no CPU path. The highest resolution setting has only been validated on a 96GB NVIDIA H20, and memory needs on smaller cards have not been measured. Training code and the paper's evaluation harness are not included.

The release uses the Hi3D Open License Agreement Version 1.0, which allows commercial use. Products or services that pass 200,000 monthly active users must notify Hi3D and apply for a separate license within 60 days, and original Apache-2.0 and MIT portions keep their terms.

Hi3D's newer commercial models are not part of this release. The company describes v1.0 as an earlier, production-tested version that developers and businesses can build on, while its Hi3D Agent product links model generation to editing, rigging and 3D-printing preparation.

SEE ALSO: ByteDance Unveils Seed3D 1.0 for Single-Image 3D Generation

XPeng says the battery it designed specifically for its IRON humanoid robot has started coming off the line in small batches. Gu Jie, a vice president of XPeng Group, said at the SAE-China Congress and Exhibition (SAECCE) on October 28 that the IRON robot's solid-state battery is now in small-batch production, while the robot itself is at the stage just before start of production, known as SOP.

All-solid-state or semi-solid-state? Today's remarks do not settle that question. As reported by Chinese media, Gu described it only as a solid-state battery, using the general Chinese term, and said XPeng had launched the industry's first solid-state battery made specifically for robots. He did not say whether it is an all-solid-state cell or a semi-solid-state one. When XPeng unveiled the new-generation IRON at its 2025 Tech Day, the company said the robot would be the first in the industry to use an all-solid-state battery, according to Yicai. In September, some English-language reports said the production version of IRON would switch to a semi-solid-state battery. XPeng has not publicly clarified which type the small-batch cells are, so Pandaily is reporting the wording as given.

Gu did share some design details. Because a humanoid has very little room for a battery pack, XPeng chose a high-energy-density design and says it reworked the battery standard around the robot. The pack uses a design that prevents thermal runaway from spreading between cells, and XPeng says it can withstand impact and nail-penetration tests.

The battery news fits into a timeline XPeng has already laid out. On September 8, the company started up its robot production line, where an IRON unit completed automated final assembly and walked off the line on its own. XPeng plans to reach mass production of the humanoid at the end of 2026 and begin deliveries in the first half of 2027. The first deployments are planned for XPeng stores, shopping malls and other commercial settings, where the robots would guide visitors and explain products, before the company expands into industrial and home use.

Energy storage is one of the hard limits on how long a humanoid can work between charges, so XPeng's choice of cell chemistry will be closely watched once the robot reaches customers.

SEE ALSO: XPeng Commissions Automated IRON Humanoid Line

Two of China's most widely used desktop AI agents shipped updates on October 28 that move in the same direction: letting the agent keep working on a computer while the user is away from it, and giving the user better ways to check in from a phone.

Doubao Work learns from demonstrations. ByteDance said the computer-use abilities of Doubao Work, its productivity agent, have been upgraded with more accurate and faster on-screen clicking. The headline feature is recording. A user can now demonstrate a task on their computer once, talking through the steps while doing it. When the recording ends, Doubao Work turns the actions and the narration into a personal skill. The next time a similar job comes up, the user picks that skill and only explains what is different this time, instead of describing every step again. Doubao Work could previously record browser actions as skills; this extends the idea to operating the whole computer.

The second change targets Macs. With a new setting that allows Doubao to work while the screen is locked, tasks that need to operate the computer keep running after the Mac locks. A user who has left the desk can send a new task to that locked Mac from their phone and see the results there, then take over again after unlocking. Doubao Work has added several features in recent weeks, including importing chats and memory from other assistants, a canvas, and parallel agents.

WorkBuddy adds live remote control. Tencent's WorkBuddy launched a remote control mode, with technical support from the Tencent EdgeOne team. Opening the mobile app connects to the user's computer and shows a live view of the desktop, so the user can watch each step the agent takes. The live view floats as a small window above the chat, and the input box still works, so an agent that goes off track can be corrected. In full-screen landscape mode, tapping the phone clicks the same spot on the computer, with a keyboard panel, common shortcuts and a virtual mouse for small targets.

Setup on the Mac requires enabling same-account remote connections and granting screen recording and accessibility permissions. Tencent says only one verified phone can connect to a computer at a time, the session runs over an encrypted channel, user input and remote screen images are not stored, and remote control uses data traffic but no tokens.

Taken together, the updates show desktop agents being designed to run in the background, with the phone as the place to supervise them.

SEE ALSO: Doubao Work Adds One-Click Import of Skills, Memory and Chat History · Tencent Launches Free WorkBuddy Education Edition

Ant Group's Ling large language model team has open-sourced Ling-3.1-flash-1019, an updated version of the Ling-3.1-flash model it released earlier this month, and published the underlying pretrained model, Ling-3.1-flash-base, at the same time. Both are available on Hugging Face under the MIT license.

Ling-3.1-flash-1019 keeps the architecture of its predecessor: a hybrid linear mixture-of-experts design with about 560 billion total parameters and roughly 25 billion active per token, and a context window of one million tokens. According to the model card, the stack combines 63 KDA linear-attention layers with 9 gated MLA layers, with 512 routed experts of which 8 are active at a time. The change in this release comes from additional post-training aimed at reasoning, tool use and multi-step execution.

On the third-party Artificial Analysis Intelligence Index v4.3.2, the new version scores 44, up three points from the previous release. The rest of the numbers are Ant's own reported results. The team says its score on AA-Briefcase, a complex knowledge-work test, rose from about 1,400 to 1,524 Elo, and GDPval-AA moved from about 1,618 to 1,644 Elo. On process automation, AutomationBench-AA went from 62% to 68%. Scientific problem solving also improved, with CritPt rising from 18% to 26% and SciCode from 54% to 57%.

The base model is the bigger addition for developers. Ant says Ling-3.1-flash-base beat DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 of 47 benchmarks across eight areas, again by its own testing, covering knowledge, reasoning, code, maths and long context. The weights are released in BF16 so teams can run continued pretraining, domain adaptation and their own post-training on top.

Ant frames the release around personal agents rather than chat. The update arrived a day after LingGuang 2.0, Ant's AI assistant app, was repositioned as a personal assistant that can keep working on tasks after a user leaves the conversation, using cross-session memory, long-running task scheduling and a dedicated cloud host. Examples Ant gives include following weather and bookings ahead of a trip, or tracking product prices against a set budget. The team also says it has trained a vision-language variant, Ling-3.1-flash-VL, which is already used inside LingGuang and will be released publicly later.

Recommended serving setups listed on the model card include SGLang and vLLM on eight-GPU nodes.

SEE ALSO: Ant Group Open-Sources Ling-3.1-flash Weights Under MIT · Ant Group's LingGuang 2.0 Enters Limited Beta as a Personal Agent

China's Jiaolong manned submersible has completed 7,000-meter-class sea trials after its second round of upgrades, Xinhua reported on October 27 from the support ship Shenhai Yihao. The trials were the acceptance test for a national key research and development program project aimed at improving Jiaolong's key operational capabilities and scientific use.

The tests were carried out during the second leg of China's 100th ocean scientific expedition. Jiaolong made three test dives in the Philippine Basin. Two went deeper than 7,000 meters, and the deepest reached 7,056.9 meters on October 2, the submersible's 465th dive. An earlier dive on September 30 reached 7,035.3 meters. All test indicators met the requirements for a 7,000-meter-class trial.

Sun Yongfu, a researcher at China's National Deep Sea Center, said the trials fully verified the reliability and safety of Jiaolong after the upgrade and after key components were replaced with domestically made parts.

Before the expedition, Jiaolong was refitted with a high-energy-density, oil-immersed lithium battery pack rated for 7,000 meters. It also received domestically produced direct-drive, low-noise DC thrusters, expanded interfaces for its emergency hydraulic system, a deep-water seawater hydraulic buoyancy adjustment system, and wide-angle, high-zoom ultra-high-definition camera systems. Taken together, the changes touch power, propulsion, hydraulics, electrical systems and observation and communications, which Sun described as a major increase in operational efficiency.

Sun said that once the upgrade and component replacement were complete, Jiaolong went straight into a series of deep-sea resource and ecological surveys in several sea areas, work he said would deepen international ocean cooperation.

The two dives past 7,000 meters were crewed by different pairs of submersible pilots, with each setting a personal depth record, and the results show that the refitted vehicle can again work routinely at the 7,000-meter class with a larger share of homegrown core components.

SEE ALSO: VY150 Deep-Sea ROV Dives 4,500 Meters in the South China Sea · CCCC Completes Design of a 10,000-Tonne Deep-Sea Survey Ship

RoboParty has released RP1, a humanoid robot it describes as a fully open-source, full-stack platform. Rather than offering only an SDK or a few APIs, the company is publishing material across the whole development chain, from mechanical and electrical design and its own joint modules to firmware, robot models, deployment and learning tools. A white paper was released on October 25.

RP1 stands 135 cm tall. The open-source version has 24 active degrees of freedom and the Pro version 30, including two in the head and seven in each arm, which suits more complex manipulation research. Head and hand modules can be swapped to match different uses such as education and research, data collection, guided tours and entertainment. RoboParty has shown the robot walking continuously, tracking human motion, recovering from pushes and getting up after falls.

The hardware centers on ROMOMO, RoboParty's integrated rotary joint module. Per the published specifications, nominal peak torque ranges from 25 to 160 newton-meters, with 48V power and CAN-FD communication. The modules pair inner-rotor motors with planetary reducers, trading low rotor inertia for fast response against stiffness, overload tolerance and cost. RP1 uses a non-coaxial hip and a parallel ankle in which two motors jointly control pitch and roll, and clamp mounting with hollow cable routing makes the joints easier to remove.

On GitHub, the RP1 overview repository links to sub-projects for mechanical source files, electronics, testing and assembly procedures, the ROMOMO joint firmware and host tools, the RP Hand dexterous hand with a C++, Python and ROS 2 SDK, full-machine testing and a controller app. The repository lists mechanical CAD, schematics, PCB layouts, Gerber files, bill of materials and assembly documentation among the hardware material, while URDF, MJCF and mesh models are available for simulation. The source article notes the exact scope of each release should be checked in the individual repositories.

On the software side, the embedded stack connects motors, inertial sensors and the battery management system through a unified device interface, running control at 500Hz and policy updates at 50Hz. An open motion retargeting tool called hhtools maps human motion onto the robot while respecting joint limits and ground contact. Under RoboParty's PartyOS development base, parts of the basic motion layer are open, including hhtools, MimicLite and UFO, a project on disturbance recovery and fall recovery. Higher layers for intent representation and agent skill orchestration remain in research.

RoboParty has opened applications for an RP1 beta program for developers and teams with real projects who are willing to file issues and feed back on the platform.

SEE ALSO: AgiBot Agile 2.0 Targets Open-World Locomotion

Southchip has launched the SC3606, a driver chip for piezoelectric fans, a form of active cooling that the company believes is one of the best fits for thin devices such as smartphones. Southchip has not named any device that will use the chip.

A piezoelectric fan has no motor, bearings, shaft or mechanical friction. It is a thin membrane vibrating at ultrasonic frequencies, and Southchip says such fans can be under 1mm thick, run at around 30dB of noise, draw power in the hundreds of milliwatts and respond in under a millisecond. The catch is that a piezo actuator cannot find its own resonant point or regulate its own power. It needs a driver that can output high voltage, lock precisely onto resonance and stay efficient.

The SC3606 comes in a 2.3mm by 2.75mm by 0.5mm FCQFN package with 22 pins. It accepts 2.1V to 5.5V input, so it can sit directly on the battery rail without an extra LDO or front-end DC-DC converter, and that range also suits the lower voltages of silicon-anode batteries. An integrated bidirectional buck-boost circuit lifts the output swing to as much as 46Vpp, or plus and minus 23V, across an output frequency range of 2kHz to 40kHz.

Resonance tracking is the core feature. Because a piezo fan's resonant point drifts with unit-to-unit variation, assembly stress, temperature and aging, the chip combines full-range sweeps and segmented approximation sweeps with real-time frequency tracking to find the best operating point between 15kHz and 40kHz on its own. Southchip says this means device makers do not need to calibrate each batch of fans on the production line or write temperature compensation.

The chip supports both sine and square wave drive. Sine drive, with low harmonic content, keeps the actuator vibrating only in its target mode and avoids the audible whine that harmonics can excite. Since a piezo element behaves like a large capacitor that charges and discharges every cycle, an energy recovery architecture returns that stored energy to the supply rail to cut system power. An I2C front end with 1.2V IO lets the chip connect directly to flagship SoCs and PMICs, so thermal management software can wake it and start the fan quickly. Protection includes over-voltage, over-temperature, under-voltage lockout and short-circuit safeguards.

Paired with Southchip's earlier SC3601 driver for piezoelectric micropump liquid cooling, the SC3606 gives device makers a choice of piezo-based cooling drivers from one supplier.

SEE ALSO: OPPO K15 Brings Active Cooling · Redmi K90 Max to Launch With Active Cooling

SoMark, a Chinese startup team focused on document intelligence, has open-sourced OXR, short for Optical Everything Recognition. The main model, OXR-1.0, has just 1.3 billion parameters and turns PDFs and document images into Markdown or JSON with layout coordinates included, data the team describes as agent-friendly. Code and weights are on GitHub, Hugging Face and ModelScope, with an online demo and a free API.

According to the official repository, OXR ranks first on three public benchmarks: 97.18 on OmniDocBench, 87.84 on olmOCR-Bench and 87.76 on DocParsingBench, with a particularly wide lead in table parsing. The team publishes full results and reproduction instructions in its benchmark documentation. In a speed test the team ran on a single RTX PRO 6000 Blackwell card with 96GB of memory, OXR parsed 100 randomly drawn OmniDocBench pages in about 101 seconds, against roughly 145 seconds for PaddleOCR-VL-1.6 and 255 seconds for MinerU2.5-Pro.

The model recognizes 13 element types, including text, titles, tables, formulas, stamps, code blocks, chemical equations and chemical structures. Tables come out as HTML or Markdown, formulas as LaTeX, chemical equations in mhchem syntax and molecular structures as SMILES strings. It also determines reading order, which matters for multi-column papers, newspapers and exam sheets where correctly recognized blocks can still be stitched together in the wrong sequence.

Two design choices stand out. First, a lightweight layout model locates regions, while the main model handles both content recognition and reading order in the same step, removing a separate ordering stage. SoMark says it is the first to make the layout module independent and pluggable, so a business can swap in its own layout model with different element categories or region boundaries and keep OXR's recognition. Second, post-training uses a recursive self-improvement loop: OXR parses unlabeled documents, external models judge the output and diagnose errors, a clustering model groups failures by cause, and targeted training data is generated for the next round without manual labeling.

The team says the repository passed 1,000 GitHub stars in its first week. An arXiv paper is listed as coming soon. For developers building retrieval-augmented generation or agent pipelines, OXR offers a small, fast open option for turning complex documents into structured input.

SEE ALSO: Tencent Open-Sources Hunyuan OCR · PaddleOCR Becomes the Most-Starred OCR Project on GitHub