EN ▾
SNAPE REP RT Front page →
ConfirmedCompute18 Nov, 05:02 UTC

China Telecom details a 285B-parameter Xing 4 model pretrained on 8,192 Huawei Ascend 910C chips

Architecture diagram of the 285B-parameter Xing 4 mixture-of-experts model
Xingchen AGI Lab, China Telecom (arXiv)

The technical report also releases an open 29B model trained entirely on Ascend with MindSpore; it reports systems measurements, not head-to-head scores against US frontier models.

Read at Xingchen AGI Lab, China Telecom (arXiv) Share

Why it matters

Most large open models are trained on Nvidia GPUs; this report documents a 285B-parameter mixture-of-experts model pretrained on Huawei Ascend 910C chips with the MindSpore framework, in collaboration with Huawei. That matters as US rules tighten Chinese developers' access to American accelerators. The report focuses on systems engineering rather than broad benchmark claims against US frontier models. Watch whether the larger Xing 4 members are released and how they score in independent tests.

Key facts

  • 285B-parameter MoE, 17B active per token, 1M-token context
  • Pretrained on 8,192 Ascend 910C cards
  • Open-sourced Xing4.0-29B-A4B: 29B total, 4B active, 256K context
  • Led by Xingchen AGI Lab, China Telecom, with Huawei and HIT (Shenzhen)
  • Up to 4.68× prefill throughput over an MLA-only baseline

Sources

More on Compute
More from Snape ReportFront page →