China Telecom Opens Local AI Model as Huawei and Qiyuan Push Enterprise and Home Robots

According to QbitAI, the model was trained entirely on Huawei Ascend 910C chips and uses the domestic MindSpore/MindFormers training framework. It has 29 billion total parameters and activates about 4 billion per inference through a mixture-of-experts architecture. It natively supports 256K context and can be extended to 512K.

QbitAI reported that the model can complete an initial review of a 200-page bid document in about 20 minutes locally, covering technical, commercial and price sections with scoring rationale. It can also turn Excel contract ledgers into a visual management dashboard without data leaving the enterprise environment. Coding and agent capabilities were strengthened, with adaptation to OpenCode, Claude Code, OpenClaw and Hermes, and compatibility with LLaMA-Factory, MindFormers, SGLang, vLLM and KTransformers.

China Telecom says 4-bit quantization cuts a 29B model's memory footprint from about 60GB under FP16 to about 15GB, a reduction of about 75%, allowing RTX 3090 and 4090 consumer cards to run it. The company said the model has been privately deployed in intelligent customer service, where user calls, identity information and business records remain within the domain, and complex business handling success rate exceeds 90%.

The model uses MLA attention, multi-token prediction and mHC manifold-constrained hyper-connections. It was trained on tens of trillions of tokens, including complex tables, structured knowledge graphs and agent behavior traces, with training stages extending from 4K to 32K, 128K and 256K sequences. Post-training uses multi-expert reinforcement learning across coding, agent and reasoning, fused through MOPD. On Ascend 910C, DVM graph-computation fusion, fine-grained recomputation and Ascend C fused operators raised training throughput by about 96% over out-of-box performance, according to the company.

QbitAI reported that Xing4.0-29B-A4B is open-sourced on GitHub, Hugging Face, Gitee, ModelScope and other platforms, with API access. As of publication, it ranked fourth on Hugging Face's model trending list. China Telecom also offers deployment through a compute-network all-in-one machine, cloud elastic expansion over an intelligent computing line, and multi-chip adaptation through FlagOS.

On September 18, Huawei released the Agentic Enterprise: Enterprise Intelligence Whitepaper at HUAWEI CONNECT 2026, QbitAI reported. The whitepaper proposes that enterprises start from business goals rather than from model selection, and introduces DIMAK, a framework covering data, AI infrastructure, model, agent and knowledge, along with an H-type twin-tower architecture that connects an intelligent system with existing IT and OT systems. DIMAK is intended to decouple the technology lifecycle from the enterprise capability lifecycle, so verified data services, knowledge assets and task methods can survive model and framework changes.

Huawei AI application chief expert Li Hongkai told QbitAI that local efficiency gains can shift bottlenecks, so companies should identify the constraint in the full business process. In an e-commerce case cited by QbitAI, a merchant integrated AI image generation, model swapping and background changes into product listing, reaching about 20 times the previous speed; the system was used by eight stores at about 20,000 yuan per store per month, but human checks were still required and the AI had not been connected to review, selection and operations. The merchant's manager, Peng, told QbitAI that self-training models without sufficient business scale would lose money.

The whitepaper also cites grid equipment evaluation and bank wealth management maturity cases. Huawei AI system scientist Yao Jun said at the launch that the last mile of the business value stream must be built by enterprises and that Huawei cannot replace that. Huawei said the whitepaper records practices in finance, automotive, pharmaceuticals, materials chemistry and power, and that it will provide reference methods, engineering frameworks and common tools while enterprises and partners handle specific implementation.

Also on September 20, Qiyuan Robotics, a personal and family consumer robotics brand under Shangwei New Materials, held a launch in Shanghai and began selling two personal robots, the Q1 and T1, both starting at 19,999 yuan, according to ifanr. The Q1 Explorer is priced at 26,999 yuan and the T1 Pro at 29,999 yuan. The first authorized experience stores will cover Shanghai, Guangzhou, Shenzhen, Hangzhou, Xiamen, Wuhan and Xi'an. The company emphasized sale means shipping, moving from technical prototypes to commercial products that require production consistency, after-sales service, quality management and software updates.

The Q1 Explorer is aimed at development and expansion, with open SDK, HDK and hardware expansion interfaces, full-stack secondary development and skill application development tools. The T1 Pro adds an NVIDIA Orin Nano chip, 360-degree omnidirectional obstacle avoidance, low-battery automatic recharge, voice summon and follow-and-carry functions, and also opens secondary development interfaces. Qiyuan disclosed a robot technology stack that includes composite materials, self-developed QDD quasi-direct-drive joints, and three AI models: PrimeMotion for motion control, PrimeGo for multimodal perception and following in humanoid and quadruped forms, and PrimeVista for visual and voice interaction. It plans a Harness-based robot agent and a PrimeMind household base model.

Qiyuan also announced PrimeFLOW for robot appearance creation, PrimeStore as a robot application store with partners including Ximalaya, QQ Music, Youdao, JD Health and Ctrip, and PrimeHub for developers to create, publish and share robot skills. Ifanr reported that more than 1,000 developers have joined the ecosystem, and Qiyuan plans to invest 1 billion yuan over five years in its Qiyuan Star Ecosystem plan, with joint laboratories or industry-education platforms with Peking University, Shanghai Jiao Tong University and Zhejiang University. The company said its production line can roll out one robot every 2.5 minutes and records assembly data for long-term traceability.

Editor's Summary China Telecom's open-source model brings a 29B-parameter agent to a single consumer GPU for private enterprise deployment, while Huawei's whitepaper sets out engineering methods for connecting AI to existing business systems. Qiyuan Robotics has begun selling personal robots from 19,999 yuan, betting that developers and early users will expand the devices' capabilities beyond initial demonstrations.