AI Agents Northeastern and Stanford Give Agents a Git History: Shepherd Lifts Pair-Coding Pass Rates From 28.8 to 54.7 Percent Coding Assistants China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack News Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek Business China's Data Regulator Will Set National Rules for Metering and Billing AI Tokens After Daily Use Tripled to 500 Trillion Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying AI Agents Northeastern and Stanford Give Agents a Git History: Shepherd Lifts Pair-Coding Pass Rates From 28.8 to 54.7 Percent Coding Assistants China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack News Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek Business China's Data Regulator Will Set National Rules for Metering and Billing AI Tokens After Daily Use Tripled to 500 Trillion Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying

China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack

China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack

A 138-author technical report from China Telecom's Xingchen AGI Lab describes Xing4, a 285B-parameter mixture-of-experts model with about 17B active parameters trained entirely on 8,192 Huawei Ascend 910C accelerators using MindSpore. The lab reports 2.67x to 3.77x decoding throughput gains from its DSpark method; the figures come from the paper and have not been independently reproduced.

China Telecom's research arm has published a technical report describing the training of a 285-billion-parameter model entirely on Huawei silicon, without a single Nvidia GPU or line of CUDA in the pipeline. The paper, titled "TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks", carries 138 authors and documents a run across 8,192 Ascend 910C accelerators.

Xing4 is a decoder-only mixture-of-experts family rather than a single checkpoint. The flagship, which has not been released, holds 285 billion total parameters with roughly 17 billion active per token and is served only through China Telecom's Xingchen TokenHub platform, which the company says processes more than one trillion tokens a day. The open member, Xing4.0-29B-A4B, shipped on Hugging Face in September under Apache 2.0 with 29 billion total and 4 billion active parameters.

The architecture reads as an adaptation to the hardware rather than a copy of it. The team combined multi-head latent attention with DeepSeek's sparse attention, added an expert-routing index mechanism it calls IndexShare, kept the first three transformer layers dense to stabilise early feature extraction, and used multi-token prediction to squeeze more output per forward pass from chips whose memory bandwidth trails Nvidia's by roughly a fifth.

The headline engineering claim is a decoding method the paper calls DSpark, which the authors credit with 2.67x to 3.77x higher decoding throughput and a 4.68x rise in prefill throughput against their baseline, plus a 1.35x reduction in step time from dropping pipeline parallelism. Those are paper figures produced by the team that built the system, and no outside lab has reproduced them.

That caveat matters because the training stack is the real subject. The run used MindSpore HyperParallel, Huawei's own distributed training framework, which means the contribution is less a model than a demonstrated toolchain: a frontier-scale job completed on domestic accelerators with domestic compilers and communication libraries. Chinese labs have published single-model results on Ascend before; this is the largest publicly documented cluster described end to end.

The commercial logic is straightforward. China Telecom is a state carrier with its own data centres, its own cloud and a token-selling business, and it has an interest in proving that domestic silicon can carry the whole stack rather than only inference. The 29B open model gives outside researchers something they can actually run, which is the only real check on the larger claims.

There are limits worth stating plainly. The flagship's weights remain closed, so its capabilities can only be tested through an API, and the report offers no head-to-head evaluation against frontier Western models on standard agentic benchmarks. Memory bandwidth, interconnect scale and the maturity of the software ecosystem are all still constraints that a single successful run does not remove.

What this means: the export-control argument has always rested on the assumption that the gap is in the toolchain, not just the chip. A 138-author engineering diary describing how a phone company's research lab closed that toolchain gap is a more useful signal than any benchmark table, because it can be copied by everyone else who has Ascend capacity.

Comments (0)

Log in to join the discussion

Log In

No comments yet