Rokid has launched ride-hailing on its AI glasses in partnership with Tencent Mobility, letting wearers book a car by voice without taking out a phone. Rokid describes it as the industry's first native ride-hailing service for AI glasses, a claim made by the company.

Once the Tencent Mobility agent is connected, a user only needs to say where they want to go. The glasses then handle the full flow: understanding the request, suggesting pickup and drop-off points, checking nearby cars, estimating fares and comparing vehicle types, placing the order and paying the fare. While the user waits, the license plate, car details and estimated arrival time are shown in their field of view, with voice reminders. According to Chinese tech outlet Leikeji, Android and iOS users need to update the Rokid AI app to version 1.15 to use the feature.

The integration works as a split between two agents. Rokid's glasses agent handles understanding the user, conversation and on-lens display. Tencent Mobility's agent provides the ride-hailing service. The two connect directly over the Agent-to-Agent (A2A) protocol. Tencent Mobility has packaged its ride-hailing functions for wearable devices as standard services: location suggestions, pickup point selection, nearby vehicle lookup, fare estimates and comparison, one-tap booking, live order status and password-free payment. The glasses call these as needed.

Rokid argues the main benefit is at the curb. Instead of repeatedly checking a phone for the car's position, plate and arrival time, users keep that information in view, which can make it easier to find the car in crowded stations, malls or roadsides. The company also says voice booking lowers the barrier for older people, children and others less comfortable with smartphone apps.

Rokid is clear that the glasses are not meant to replace the phone. For payment issues, changing the destination or cancelling an order, the phone remains the more reliable fallback. In the company's view, glasses should start requests, give live reminders and handle interaction with the surroundings, while the phone takes care of complex confirmations, payment and exceptions.

The companies have not said which glasses models or cities are supported.

SEE ALSO: JD Technology and Rokid Unveil JoyGlance · Sharge Launches loomys L1 AI Glasses

Unitree Robotics has released UniUni, a swarm control system for choreographing and supervising groups of its robots. The software is free for all Unitree users. At launch it supports the company's H2 and G1 humanoid robots.

UniUni is a desktop application for Windows and macOS. The versions listed at launch are 1.1.14 for Windows and 1.1.15 for macOS. Unitree says the system is built to plan the movements and positions of multiple robots quickly, while monitoring device status and handling safety stops in real time.

Path planning happens in both 2D and 3D views, where users draw the trajectory each robot should follow. A drag-and-drop timeline lets them sequence moves across the group, and an AI-assisted tool arranges whole-body motions and syncs them to music. For positioning, users can import point-cloud maps, with robots localizing themselves through laser SLAM.

Setup is designed to be light. Robots on the same local network are discovered and paired automatically. During a run, operators get live status for each machine and a one-tap emergency stop for the whole group.

The release turns group-control tooling into a product that customers can run themselves. Coordinated humanoid performances have become a common way for Chinese robot makers to show their machines in public, and they typically require careful planning of timing, spacing and fallback behavior. A shared timeline and map, plus central monitoring and a single stop control, address those needs directly.

Unitree has not disclosed the maximum number of robots UniUni can coordinate at once, and has not said which other models will be supported after the H2 and G1. It has also not published technical details on network requirements or positioning accuracy.

UniUni adds to a software layer Unitree has been building around its hardware. The company has recently opened a project page and partial weights for its UnifoLM-WLA-1.0 humanoid foundation model, and refreshed the G1 line with the G1+.

SEE ALSO: Unitree Unveils G1+ Humanoid With Six Upgrades · Unitree Opens Partial Weights for UnifoLM-WLA-1.0

UBTECH has published details of Thinker-WM, a world model developed by its Thinker team that encodes RGB video, metric depth and robot actions in a single shared representation. A technical paper is available on arXiv. The company says the model ranks first on three embodied world model benchmarks. Those rankings are UBTECH's own claims, based on results available when it made the announcement.

By UBTECH's account, Thinker-WM scored 0.740 overall on WorldOlympiad, a benchmark built on real robot interaction data that covers more than ten robot bodies and several camera setups. It also posted the top single scores for 3D consistency (0.531) and interaction fidelity (0.765). On EWMBench it reported a total of 4.66, ahead of models including Qwen-RobotWorld and Sora 2, with leading results on trajectory-matching and dynamics metrics. On WorldArena 1.0 Track-1, a simulation benchmark, the company says it led across 12 evaluation dimensions.

The design argument is that many world models are stitched together: video generation handles pixels, a policy network handles actions and a separate module predicts geometry. UBTECH contends that this splits signals that are physically coupled. Thinker-WM instead maps each modality through its own encoder into a common latent space and models them jointly in a shared Unified Dynamics Transformer before decoding video, depth maps and executable actions. Depth is treated as a native output rather than an auxiliary signal. In single-task ablations on RoboTwin, adding depth raised manipulation success by 15 percentage points over a baseline without it.

Because the modalities share one set of weights, swapping which inputs are given and which are generated lets the same model run six modes, including joint video-action generation, future prediction, policy inference, video-to-action and inverse dynamics.

Training data totals roughly 30,000 hours: about 20,000 hours of first-person human video from homes, retail and factories, about 10,000 hours of real robot data from several robot types, and about 1,000 hours of simulation. All robot actions are expressed as end-effector poses, with a body-specific tag so different machines share one generation space. A data pipeline called Data Factory handles sampling, layered quality checks and depth-video alignment.

To test whether the shared space truly couples modalities, the team trained a small motion probe on frozen visual features to recover end-effector displacement, which UBTECH presents as evidence of cross-modal coupling.

SEE ALSO: WorldArena 2.0 at IROS 2026 · UBTECH and FAW-Volkswagen Take Humanoid Robots Into Factory Logistics

Tencent Hunyuan has released version 2.0 of Hy Memory, its memory framework for AI agents, and open-sourced all three layers of the stack: a dedicated memory model, the framework itself and a local memory app called MemHub. The layers can be used separately or combined. Tencent says Hy Memory is already used by internal teams in finance, news, knowledge management, office agents, social entertainment and games.

The framework handles storage, updating and evolution of three memory types: chat, working and code. Developers connect it to agents through an SDK, published on PyPI as hy-memory 2.0.2. Working memory and code memory are new in 2.0. They record a task's goal, decisions already made, intermediate results and reusable lessons along with the conditions under which those lessons apply. Before resuming a long-running task, the system first works out which task is in progress, then retrieves relevant decisions, still-valid experience and prior work history. Chat memory now groups scattered facts into events with time, place and people, merging duplicates and resolving conflicts before passing context to the model.

Tencent reports first place against open-source frameworks including Mem0, MemOS and EverOS on four benchmarks. By its figures, working memory scored an average 55.31% on MemoryArena travel-planning and shopping tasks. Code memory reached 64.2% question-answering accuracy and 35.6% task resolution on CodeMemoryBench, and 28.3% task resolution on SWE-ContextBench. Chat memory hit 72.6% accuracy on LoCoMo while cutting tokens per question from 1,159 to 703, a 39.3% reduction.

The model layer is Hy-Memory-30B-A3B, a mixture-of-experts model with about 3 billion active parameters trained for memory extraction and conflict handling. Its technical report evaluates it on a fixed 20% subsample of the benchmarks across four memory systems. On that subsample it posts a macro average of 55.49, slightly above Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64. Weights are on Hugging Face.

MemHub is a local-first app with a command-line tool and a Mac client. Once a user authorizes connected tools, it imports their past sessions into a single local memory store and keeps syncing new ones. Tencent says it currently plugs into 10 AI tools, including WorkBuddy, Cursor, Codex, Qwen Code, OpenCode and DeepSeek Harness.

The GitHub repository's license file states that Hy-Memory is released under Apache 2.0.

SEE ALSO: Tencent Hunyuan Releases Hy4 Preview · Doubao Work Adds One-Click Import of Skills, Memory and Chat History

DeepSeek has added two standalone tools to its API platform: a Search endpoint that queries the web and returns structured results, and a Fetch endpoint that retrieves a single page and returns its content as Markdown. According to the company's Chinese-language API documentation, the tools are meant to give agent and retrieval-augmented generation (RAG) applications access to live web content without a separate third-party search provider.

Both are raw JSON APIs that use the same Bearer authentication as the rest of the DeepSeek API. Search is served at api.deepseek.com/v1/tools/search and Fetch at api.deepseek.com/v1/tools/fetch.

A Search request takes a query of up to 128 characters and can return between 1 and 50 results, with 10 as the default. Developers can pass the user's original question to help the service read intent, supply an IP address to favor localized results, and ask for full page text. Queries accept operators including site: and -site: to include or exclude domains, quoted exact phrases, lang: to restrict languages, and after: and before: to bound publication dates. Each result carries a title, URL, a short snippet of roughly 100 to 200 characters for display, a longer summary of roughly 600 to 800 characters intended for models to read, an optional full-text field, a publication date, a rerank score from 0 to 3 and an authority rating from 0 to 4.

Fetch takes a single HTTP or HTTPS URL and returns the page title plus Markdown body. DeepSeek notes that results depend on the target site's availability and access policies, and that blocked or unextractable pages return an error.

Pricing is per request rather than per token or per result. Search and Fetch each cost 5 yuan per 1,000 successful requests. Only calls that return HTTP 200 with content are billed: a Search must return a non-empty result list and a Fetch must return page text. Failed requests, empty searches and timeouts are free, and the number of results requested does not change the price.

Searches triggered through the Web Search tool in Claude Code or the web_search feature in Codex will not incur Search fees for the first two weeks after the Search API launch. DeepSeek says standard pricing will apply afterwards and that it will announce the exact billing start date separately.

Rate limits are set per account, not per key. Search and Fetch are each limited to 5 queries per second, tracked independently of each other and of model concurrency limits. Fetch also applies per-domain limits, and exceeding either cap returns HTTP 429.

SEE ALSO: OpenCode Makes DeepSeek V4.1 Flash Free to Use · DeepSeek Open-Sources Harness Agent Runtime

Tencent's Toast App, the vibe-coding product under its MyApp store brand, now lets Pro users import external Android APKs and H5 source code, then keep editing inside Toast before filing for software copyright, filing for record and submitting to MyApp. The change widens Toast from apps born inside the tool to startups that already built elsewhere.

According to IT Home, APK import is meant for teams that finished development, debugging and packaging in other tools. They upload the package, fill in listing details such as description, icon and screenshots, and enter the same listing path Toast already offered for in-app creations. H5 import keeps room to iterate: after upload, developers can preview, continue building in Toast and let the platform auto-capture a product blurb and screenshots for store materials.

Toast opened an end-to-end listing flow in August that covered copyright application, app filing, compliance paperwork and MyApp review. That path previously served apps generated on Toast itself. External import extends the same pipeline to multi-channel prototypes, which matters for solo builders and one-person companies that do not want to rebuild an MVP from a prompt.

Access is limited for now. IT Home and Ifanr both note that APK and H5 import sit inside the advanced Pro benefits package, not the free tier. Toast still handles creation and listing prep; MyApp handles review, distribution and ongoing store operations.

For Tencent, the move turns Toast into a broader on-ramp for Android listings rather than only a generation sandbox. For developers, it is an incremental feature on top of the May product launch: bring an existing APK or H5 project, polish it, and walk the compliance path into MyApp without starting over.

SEE ALSO: Tencent's Toast App Lets Anyone Build Native Mobile Apps with Natural Language

4Paradigm and RiseUnion have open-sourced ModelSphere, a production-oriented AI deployment platform that unifies compute management, model serving and runtime tuning under the Apache 2.0 license. The code is published on GitHub at modelsphere/modelsphere, which lists an Apache-2.0 license and was pushed again on the launch day.

The companies say ModelSphere sits above engines such as vLLM and SGLang: those stacks squeeze more work from a single server, while ModelSphere handles cluster routing, a shared KV cache pool, quality-of-service autoscaling, hang recovery and online AutoTune. Built-in features listed in the joint announcement include a graphical deploy panel, elastic scale-out, prefill-decode split serving, three-level KV caching, high-availability load balancing and automatic resume after a service break.

Heterogeneous chips are a stated focus. Through an EngineX inference layer, ModelSphere is said to schedule Huawei Ascend, Iluvatar, Moore Threads and Hygon DCU resources in one fleet so teams do not rewrite low-level code for each accelerator. The project already lists adaptations for DeepSeek-V4 and V4-Flash, Qwen, Kimi, MiniMax and GLM, with a goal of production-ready configs soon after new open models appear.

On performance, the announcement claims AutoTune and hardware-aware balancing can raise peak throughput by about 285% versus an internal baseline, described as roughly a fourfold jump. That number is the companies' own claim, not an independent audit, and should be read as a marketing figure until outsiders reproduce it.

In the stack diagram the partners draw, single-node engines stay responsible for kernels and batching, while ModelSphere owns request routing that notices cache locality, a unified KV pool across replicas, SLO-driven scale decisions and operational monitoring. The pitch is that enterprises which already own mixed accelerators can stand up stable LLM APIs without rebuilding a private control plane from scratch.

The release is free for enterprise use under Apache 2.0 and is framed as infrastructure software, not a financing or earnings story.

SEE ALSO: OpenCode Adds Free DeepSeek V4.1 Flash With a 200K Context Window

WanNa Robotics has introduced NOVA21, a 21-degree-of-freedom dexterous hand aimed at research labs and robot developers, at the 2026 Hong Kong Autumn Electronics Fair. The company calls the debut a global first; that claim is its own marketing language and has not been independently verified.

NOVA21 uses a human-like layout at about 195 by 105 by 71 millimeters. The thumb has five active degrees of freedom and each of the other four fingers has four, all driven by rotary direct-drive actuators. WanNa says fingertip force reaches at least 20 newtons per finger and four-finger grip force at least 200 newtons, with force control and backdrivability for contact experiments.

Speed and accuracy figures are also public. Finger curl under no load is listed at 520 degrees per second, with fingertip repeatability of plus or minus 0.2 millimeters. Dual encoders on the drive modules are rated at 0.2 to 0.5 degrees of angular repeatability.

For touch, the hand carries 16 force sensors that combine fingertip and finger-pad three-axis sensing with a palm normal-force array. Normal force range is given as 0 to 25 newtons, tangential as plus or minus 10 newtons, and tactile resolution as 0.1 newtons. Interfaces include RS485 or CAN-FD on a two-pin connector and EtherCAT or TCP/IP on a four-pin connector, plus OTA updates and a quick-swap wrist.

WanNa positions NOVA21 for multi-finger coordination studies, industrial validation tasks such as tool pickup and connector insertion, and service-style pick-and-place across varied packaging. The firm says it designs the miniature actuators, structure and control stack together so force and motion can be tuned as one system.

SEE ALSO: Inspire Robots Launches RH37D1, a Direct-Drive Three-Finger Hand That Lifts Up to 43kg

AgiBot has begun rolling out its YuanSheng ecosystem program into Chinese universities, saying a first group of more than 20 schools will co-build teaching and research platforms for embodied AI. Named partners include Peking University, Tsinghua University, Zhejiang University, Shanghai Jiao Tong University, Xi'an Jiaotong University, Southeast University, Sun Yat-sen University and South China University of Technology.

The company plans to place more than 1,000 real robots on campus, spanning wheeled Genie units, compact Lingxi humanoids, full-size Expedition humanoids and quadrupeds from its AgiQuad line. The goal, AgiBot says, is to close the gap between classroom theory and scarce research hardware so students can train models, debug algorithms and run tasks on physical machines.

Software comes with the hardware. Universities get access to AIMA, AgiBot's embodied development stack that includes LinkSoul for agent applications, LinkCraft for robot authoring and Genie Studio for end-to-end task workflows. Together, the machine fleet and software membership rights are valued at about 200 million yuan under the company's own accounting of product and license value, not as an external financing round.

YuanSheng itself was framed earlier this year as a multi-year ecosystem effort. AgiBot says the program has already connected with 237 research institutions and more than 500 partners, and counts over 10,000 application developers plus more than 70,000 subscribed developers. The university track adds course co-design, joint research and hands-on training to that partner network.

The move is an education and talent play rather than a single product launch: AgiBot is packing bodies, software seats and curriculum hooks into one package so labs can run embodied AI projects without waiting for scarce robot inventory.

SEE ALSO: AGIBOT Unveils Open-Source Simulation Platform Genie Sim 3.0 at CES 2026

Kuaishou's Kling AI says Kling 4.0 is now fully available on klingai.com, moving the model from preview access to a general rollout. The update focuses on longer native clips, richer multi-reference control and audio that ships with the picture, rather than rehashing the September capability preview.

Creators can generate a single native clip of up to 30 seconds and set as many as 10 keyframes to steer character state, scene changes and plot beats. A generation job can combine up to 10 images, five video clips and seven subjects so references for character, motion, camera and sound work together. The product also supports multi-shot continuation that can extend a sequence toward about two minutes.

Audio is part of the full launch package. Kling 4.0 supports dual-channel stereo and lip sync across multiple languages and regional accents, including Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, plus several Chinese dialects and English accent variants. Prompt input can run to about 8,000 tokens, with the system able to expand and polish instructions in a director-style pass.

Kling AI says 4K and 1080p 10-bit HDR output, plus an enterprise API, are coming soon, and that Kling 4.0 capabilities are already wired into MCP and CLI workflows for broader tooling. The company also highlights stronger camera moves, style controls and edits that add, change or remove subjects and backgrounds inside an existing clip.

The September preview already outlined the 30-second native length and multi-reference ideas; this release is the moment those features open to all users on the public site, with audio, MCP/CLI hooks and the near-term HDR and API dates as the main new commercial details.

SEE ALSO: Kling AI Previews Kling 4.0 With 30-Second Native Clips and Multi-Reference Control

Axera has unveiled the AX8860, a 6-nanometer edge AI chip series aimed at running large models on factory floors, warehouses and robots rather than in distant data centers. The company presented the line at CIIF 2026 in Shanghai and framed it as a compute base for physical AI that must sense, decide and act on site.

According to Axera's official release, the AX8860 delivers up to 256 TOPS of mixed-precision compute with native FP4 support, up to 273 GB/s of memory bandwidth and as much as 128 GB of LPDDR5X. It supports Axera's AXLink multi-chip cascading and PCIe 5.0 for scaling, plus 64-channel video decode and 16-channel encode for multi-camera vision workloads.

Around the chip, Axera also showed the Yuanxi A256S module and related inference cards and servers. The software stack is meant to look familiar to cloud developers: the company says the series supports vLLM, SGLang and PyTorch and offers an OpenAI API-compatible service interface, with tools covering quantization, conversion, compilation, profiling and debugging.

Partner channel claims have drawn attention beyond the official brief. Hardware maker Sipeed, cited by Wccftech, said it measured about 25% faster prefill and 30% faster decode than an NVIDIA DGX Spark on a Qwen 27B run, and described pricing as roughly half that of a DGX Spark. Those figures are Sipeed's claims, not a third-party lab benchmark, and are not stated in Axera's own CIIF announcement.

Wccftech also reported module designs around 40 W, single- and four-chip PCIe cards, and a roadmap toward higher-TOPS follow-ons. Axera says four-chip cascading can reach about 1,024 TOPS. The company positions the line for smart manufacturing, logistics and robotics, arguing that inference should happen where events occur rather than round-tripping to the cloud.

SEE ALSO: Axera Debuts 5nm M9 ADAS SoC Series With Up to About 720 TOPS

Hesai, a lidar company known for automotive sensors, has launched Kosmo, a 3D world data engine aimed at one of Physical AI's persistent bottlenecks: high-quality real-world data at scale. The stack pairs Kosmo One, which Hesai calls the world's first 8K ultra HD AI spatial camera, with the Kosmo World cloud platform to turn physical spaces into high-fidelity 3D digital environments for robot training, world-model development and immersive content.

Kosmo One combines high-resolution lidar, a one-inch imaging sensor and a high-precision inertial measurement unit (IMU). Hesai says the sensors stay synchronized at the microsecond level so the device captures geometry, color, texture and camera pose together. Reconstruction uses 3D Gaussian Splatting (3DGS), which the company says can reach millimeter-level geometric accuracy and reproduce materials, lighting and shadows. A single operator can scan a typical 100-square-meter indoor space in about 15 minutes, according to Hesai, and Kosmo World then builds 3D models at speeds the company puts at up to roughly five times those of comparable products. Hesai notes that "8K" refers to maximum capture resolution, while exported files top out at 6K.

Kosmo World is powered by Hesai's Kosmo Spatial Core (KSC). In one reconstruction pass it produces 3DGS, mesh and depth data in a single consistent scene. Generative tools can fill unscanned areas and rebuild materials, while semantic segmentation labels individual objects. Views can also be adapted to a robot's camera characteristics to narrow the gap between simulation and the real world.

A Dairy Queen ice cream store deployment illustrates the pitch. Hesai says building a 3D training environment there previously took about 300 hours of manual modeling; with Kosmo, scanning and reconstruction took 15 minutes. In that setup, the company says a Sharpa robot reached a 99% task success rate and cut its failure rate to one-tenth of the previous level. Separately, Hesai says a single captured space can automatically yield more than 10,000 training trajectories, delivering more than 100 times as much high-quality training data for the same cost. Those figures are Hesai's own claims.

Beyond robotics, Kosmo plugs into an open ecosystem that includes NVIDIA Isaac Sim and Isaac Lab, Unreal Engine, Pixar's RenderMan and Houdini. Hesai announced partnerships with embodied AI firms Galbot and Dexmal, is working with Amap on Flying Street View reconstructions already live for venues such as the Gravity Art Museum and PIFO Gallery, and is collaborating with Yuemo Studio on AI-assisted filmmaking and with TRIC on a digital preservation project for Du Fu Thatched Cottage in Chengdu.

SEE ALSO: Hesai's Three Co-Founders Launch New General Robotics Startup Sharpa · Sharpa Unveils D01 Tactile Robot, W02 Dexterous Hand and AE01 Haptic Glove at IROS 2026 · Galbot Prices ET1 Bipedal Humanoid From 79,000 Yuan, With Watch-and-Learn Skills on Upper Tiers

Beijing-based commercial launch startup Welight says a preburner for its Huaguang-1 rocket engine has completed a long-duration hot-fire test. Huaguang-1 is a full-flow staged combustion engine that burns liquid oxygen and methane.

The test follows earlier short-duration ignition runs, which showed that the preburner could ignite and sustain combustion briefly. The company says the component has now moved from "it lights" to "it holds steady," and calls the result a key milestone before a full-engine hot-fire test.

A preburner burns a small amount of propellant to produce hot, high-pressure gas that drives the turbopumps feeding the main combustion chamber. A full-flow staged combustion engine uses two of them: a fuel-rich preburner and an oxidizer-rich preburner, with both propellant streams passing through them before meeting in the main chamber. This approach can lower turbine temperatures and thermal loads for the same power, which suits high performance and reuse. It is also hard to engineer, because the two preburners, two turbopumps and precise timing and control are tightly coupled.

According to the company, the earlier short tests mainly checked ignition and startup. The long-duration run focused on behavior under sustained combustion and on how the thermal environment affects the component. The test collected data for combustion stability analysis, thermal protection design and structural reliability checks. The preburner work has progressed from injector water-flow tests to short ignition and now to long-duration firing.

The company has not published burn duration, chamber pressure or other test figures. Welight says it will continue analyzing the data and moving forward with component development and system integration toward a full-engine hot-fire test, without giving a date.

Huaguang-1 is intended to power the company's Weiguang-1 medium-lift reusable launch vehicle, which is designed to use a cluster of the engines on its first stage. Full-flow staged combustion remains rare in operational rocket engines worldwide, so a working component is an early but meaningful step.

SEE ALSO: Cosmoleap's 150-Ton Qingyu-11 Methane Engine Clears Full-Condition Valve Tests Ahead of Hot-Fire

OpenCode, the open-source AI coding agent, says DeepSeek V4.1 Flash is now free to use on its platform. Calls to the model inside OpenCode no longer draw down plan credits.

The free version has a smaller context window. According to IT Home's tests, it supports 200K tokens, down from the 1 million tokens available in the full version of the model.

The move builds on what OpenCode calls Operation Cheepseek. In September the company temporarily raised the allowance for DeepSeek V4.1 Flash in its OpenCode Go plan from $15 to $60, then lifted it to $80 on October 12. Free access now replaces the allowance for this model.

The change applies only within OpenCode. DeepSeek's own platform still has no free API tier and continues to bill V4.1 Flash on a prepaid, pay-as-you-go basis. DeepSeek released V4.1 Flash on September 10.

Users say speed is uneven. Several developers reported wide swings in output speed. One user said they saw peaks of about 490 tokens per second on the Ollama cloud but drops to 50 or 60 tokens per second from midday into the afternoon. Within the OpenCode Go plan, some users said speed fell below 100 tokens per second. Community tests also suggest speed drops as context grows, with reports of under 100 tokens per second past 100K tokens and about 50 tokens per second near the 200K limit. These are user reports, not official benchmarks.

OpenCode has not said how long the free offer will last. For developers, the trade-off is free access to a capable coding model in exchange for a shorter context window and less predictable speed than the paid route.

SEE ALSO: DeepSeek Ships V4.1 Flash GA With Causal-Encoder-Decoder MoE as V4 Pro Retires

Several domestic computing platforms moved quickly after StepFun opened the weights of Step 5 Preview on October 15. Huawei, MetaX, Biren and Moore Threads each announced Day-0 support for the model the same day, meaning it could be loaded and served on their hardware from release.

Huawei said its Ascend platform supported the model through the vLLM Ascend inference engine and had built an optimization package around the model's architecture and the Atlas 950 SuperPoD hardware. The work spans fused operators written in Ascend C and Triton, multi-stream parallelism, reduced compute bubbles, MTP speculative decoding and NPU graph execution.

Huawei also said the first 1,024-card Atlas 950 SuperPoD liquid-cooled system was lit up on October 15 alongside the Step 5 Preview release. The SuperPoD product line was announced earlier; what is new is that the first full 1,024-card unit is now running, with Step 5 Preview as its launch workload. Huawei describes the system as delivering 1 EFLOPS at FP8 and 2 EFLOPS at FP4, with 256TB of globally unified memory addressing and interconnect latency as low as 3 microseconds, and says its performance is on par with advanced computing systems elsewhere. Those figures are Huawei's own.

MetaX said its C-series GPUs completed Day-0 adaptation using its in-house MXMACA software stack. The company says it has delivered Day-0 adaptation for 43 flagship models since December 2025, including models from Zhipu, Alibaba's Qwen, MiniMax, DeepSeek, StepFun and Tencent Hunyuan.

Biren said it adapted the model on its general-purpose GPUs using vLLM and its BIRENSUPA software stack, bringing up weight loading, service startup and text generation and completing function and accuracy checks. The company said its SUPACODE multi-agent platform automatically parsed and adapted key modules of the model, including DSA, SummaryCache and the MoE layers.

Moore Threads said its MTT S5000 training and inference card and MUSA software stack ran the model on day one, with tuning across hardware, software and the inference framework.

Same-day support from multiple vendors is becoming a routine part of major open-weight releases in China, giving developers more choice of hardware for deployment from launch. For details of the model itself, see our coverage of the open-weights release.

SEE ALSO: StepFun Releases Open Weights for Step 5 Preview, Publishes Technical Blog on Its Architecture · Huawei to Publicly Demonstrate Atlas 950 SuperPoD AI Computing Hardware at WAIC 2026

Leju Robotics and auto parts maker Zhaofeng have put 15 humanoid robots into routine operation on a parts loading line at Zhaofeng's No. 3 plant in Xiaoshan, Hangzhou. The two companies presented the joint embodied AI loading solution at the factory on October 15.

The robots pick small components, such as wheel hub bearings with inner diameters of 80 to 120 millimeters, from wire mesh bins and place them at preset positions on machining stations. According to a Zhaofeng representative quoted by Securities Times, the robots cover 15 workstations and more than 10 types of parts, with a loading success rate holding at 98%. The deployment has run for more than two months, logging about 1,000 hours and helping produce more than 60,000 parts.

Speed is still the gap. Securities Times reported that a single loading cycle takes under 40 seconds, against about 15 seconds for a skilled worker at the same station. Jiemian put the robot cycle at about 35 seconds. Zhaofeng's chairman told reporters that fixed robot arms at the plant reach 10 to 15 seconds on some steps, and that humanoids would need to get to about 15 seconds before wider adoption makes sense. He expects that to take six months to a year.

Why humanoids in a plant that already has fixed automation? Zhaofeng's general manager said some overseas aftermarket orders run only 250 to 300 sets, with many product types and frequent changeovers. Humanoids fill in the moving, loading and adjustment work between automated steps rather than replacing the line.

Leju says it split the loading skill into more than 30 basic modules that can be combined without low-level coding, cutting development from months to weeks. A multi-robot scheduling platform pushes the same skill to many units, and safety modules detect dropped parts, collisions and overspeed. Leju CEO Chang Lin said the next priority is speed, including making better use of both arms to pick more parts at once.

The fleet also collects data as it works, more than 40,000 records a month at current rates, according to the companies. They are donating an auto parts loading dataset from real robot operation to the OpenLET community under the OpenAtom Foundation. Jiemian noted that moving the setup to other factories will still require adapting to new parts, equipment interfaces and safety rules.

SEE ALSO: UBTECH and FAW-Volkswagen Take Humanoid Robots From Quality Inspection Into Factory Logistics · Inovance Debuts InoBot W1, a Liquid-Cooled Wheeled Humanoid Built for Factory Shifts

Xiaomi has released MiDashengLM-Spatial, a new member of its Dasheng audio model family that is designed to understand not only what a sound is, but also where it comes from. The company published the full model weights and inference code under the Apache 2.0 license, which permits both research and commercial use.

The model is built on MiDashengLM-7B, the sound-understanding large model Xiaomi open-sourced earlier, and adds a spatial audio encoder called Spatial-Dasheng. According to the technical report posted to arXiv, a hierarchical semantic-to-spatial conditioning module captures semantic and spatial information separately and then merges them. The result is a single end-to-end model that can describe speech, environmental sounds and music while also reasoning about a source's direction, distance and movement. The Hugging Face listing puts the model at roughly 8 billion parameters.

Spatial cues depend on multi-channel input. The model card says the model was trained on binaural, two-channel audio and that mono input still runs but carries no directional information. Xiaomi says the model keeps its general audio abilities, and the research team reports that performance on basic audio tasks remains comparable to large audio-language models of around 8 billion parameters.

Xiaomi describes MiDashengLM-Spatial as the industry's first open-source end-to-end spatial audio understanding model. The paper makes a similar claim, qualified with the phrase "to our knowledge." That claim comes from Xiaomi and its researchers and has not been independently assessed.

Over the past two years Xiaomi has open-sourced the Xiaomi Dasheng sound foundation model and MiDashengLM-7B. The company says technology from the Dasheng series already powers more than 30 applications across its smart home and in-car cockpit products, covering speech, ambient sound and music perception. That figure refers to the existing series, not to the new model.

For MiDashengLM-Spatial itself, Xiaomi says it will be rolled out gradually in smart home and car cockpit scenarios, where knowing where a voice or alarm is coming from could make assistants respond more precisely. The company has not given a timeline or named specific products.

The release includes a project demo page, the arXiv technical report, evaluation code on GitHub and model weights on Hugging Face.

SEE ALSO: Xiaomi Open-Sources MiMo-V2.6 Pro and Flash Weights Plus RL Stack

StepFun released the weights of its Step 5 Preview model on October 15, keeping the open-sourcing date it announced when the model launched in September. The release includes BF16 and FP8 checkpoints on Hugging Face and ModelScope, a technical blog describing how the model was built, and supporting open-source developer tools.

Step 5 Preview is a sparse mixture-of-experts model designed for long-horizon agent tasks. It has 600B total parameters, of which 27B are active per token, and supports a context window of up to 1 million tokens. StepFun says it performs well in coding, software engineering, professional knowledge work and finance. The company also says the model currently ranks second worldwide on OpenRouter's weekly and daily usage charts.

Developers can serve the model through an OpenAI-compatible API with vLLM or SGLang, or run offline inference with Transformers. According to the model card, the FP8 checkpoint fits on a single node of four B300 GPUs or eight H200 GPUs, while the BF16 checkpoint needs eight B300 GPUs or two eight-GPU H200 nodes. StepFun notes that the full BF16 sharded workflow for Transformers has not yet been validated end to end.

The weights are not under a standard open-source license. They are released under the StepFun Model License 1.0, which the company describes as Apache-2.0-style but with additional attribution and commercial-license conditions.

The technical blog explains how StepFun co-designed the architecture and training for long agent runs. The model uses a narrow but deep design with 92 layers, which the company says gives more computation steps for implicit multi-hop reasoning while keeping parameter count and compute in check. To cut the cost of attention over a million-token context, it uses Sparse GQA, which combines block-level indexing with sparse computation. Indexing in blocks of 8 tokens reduces the cost of indexing and top-k selection to one-eighth.

For training stability in such a deep network, StepFun keeps key states and parameters in FP32 precision and applies fine-grained gradient clipping to curb activation growth and gradient spikes. Reinforcement learning ran across about a million verifiable environments covering software engineering, scientific reasoning, machine learning research and professional knowledge work. StepFun also built context compression into end-to-end policy training, so the agent learns to summarize task state and continue working, and paired it with finer-grained reward attribution for long trajectories.

SEE ALSO: StepFun Launches Step 5 Preview: 600B Sparse MoE, 1M Context, Weights Open Oct 15 · StepFun Step 5 Preview: ~600B MoE Long-Horizon Agent; Open Weights Due Oct 15 · StepFun's Step 5 Preview Comes to Tencent WorkBuddy, Free to Use Exclusively Until October 25

Alibaba's workplace AI agent QwenWork has released an e-commerce edition and opened it to invited testers on October 14. Users switch QwenWork into an e-commerce operations mode to try it. It is the first industry-specific version of the product.

The edition targets the daily work of merchants who sell on several platforms at once. Its main features are a cross-platform operations dashboard, trend insight and selling-point analysis, batch generation of marketing content, and cross-platform return-on-investment optimization. It also keeps QwenWork's general features such as content generation and data analysis.

QwenWork CEO Chen Yusen said the edition is positioned as an independent, third-party e-commerce agent open to sellers on every platform, and that merchant tools and industry software from any platform are welcome to connect. QwenWork already links to a number of mainstream merchant tools, including Alibaba's own merchant analytics and seller workbench tools, Alimama's advertising tools, Youzan, FastMoss and LinkFox. The company says it will keep opening up to more tools.

According to LatePost, preparation began in June and intensive development in September, with close to 100 people involved. The e-commerce team contributes industry skills and merchant insight, and the QwenWork team provides general agent capabilities. The current homepage has five sections: a seasonal campaign area, business analysis, product management, content production and marketing.

LatePost also noted the edition's current limits. For now it works more like a cross-platform analysis and content production workbench. It cannot adjust a merchant's ad placement plans or spend money on promotion directly, and it is not a fully autonomous agent that can take over running a store. Alibaba has said more proactive features are planned, such as automatically monitoring inventory and advertising for anomalies.

The release puts QwenWork alongside other Chinese workplace agents moving into industry tasks. LatePost said ByteDance keeps industry skills inside the general Doubao Work product, while Tencent focuses on letting partners build apps on WorkBuddy.

SEE ALSO: Alibaba Opens Sign-Ups for Qwen Go!

Inspire Robots, a Beijing-based maker of dexterous robot hands and micro actuators, launched the RH37D1 series at the 2026 China International Industry Fair in Shanghai. The new product is a three-finger hand built for heavy industrial grasping rather than for imitating the human hand.

The RH37D1 uses a fully direct-drive architecture and has seven active degrees of freedom. Its maximum pulling load is 43kg and its fingertip force is 40N. According to the company, it combines compliant control with several grasping modes and is aimed at industrial automation and smart manufacturing.

Three fingers sit between the two-finger grippers common in factories and five-finger humanoid hands. They can still pinch with opposing fingers while handling a wider range of shapes, and need fewer motors and sensors than a full five-finger design. That keeps cost and complexity down, a trade-off many industrial users care about more than human-like dexterity.

The launch comes about two weeks after Inspire Robots debuted the RH5MK1 at IROS 2026 in Pittsburgh. That hand is the opposite type of product: human-sized and fully direct-drive, with 22 active degrees of freedom, five independently moving fingers, a 5Hz opening and closing frequency and 6D fingertip tactile sensors with a force resolution of 0.005N.

At the Shanghai fair, the company also showed its E2, F1, G2, H1, J1 and K1 hand series, which differ in degrees of freedom, drive method and sensing. Its industrial-grade RH56E2, shown at IROS, has a fingertip grip force of 30N and a whole-hand load of 15kg. A line of micro servo electric cylinders, which pack a coreless motor, planetary gearbox, roller screw and sensors into one compact unit, was also on display. Live demos included arm-and-hand grasping, gesture recognition with visual tracking and a dexterous hand turning a valve.

Inspire Robots has not yet published the RH37D1's size, weight or price.

SEE ALSO: Unitree Launches Dex5-S 22-DoF Human-Sized Dexterous Hand · China's Dexterous Hand New Top 4