huggingface.co web signal

TeleChat Xing4 Trains 285B-Param Agentic MoE Entirely on 8,192 Ascend 910C NPUs With MindSpore

Summary

A China Telecom-led technical report posted Nov 18 details Xing4, a frontier agentic LLM series trained entirely on 8,192 Ascend 910C NPUs using the MindSpore HyperParallel stack — no GPUs, no CUDA. The paper introduces a 29B-A4B open-weight model (29B total, 4B active) with 256K-1M context, and a 285B-A17B flagship, combining Multi-head Latent Attention, DeepSeek Sparse Attention and IndexShare, plus DSpark speculative decoding that claims 2.67-3.77x throughput. Reported gains include 4.68x prefill throughput over the MLA baseline and a 1.35x step-time speedup from removing pipeline parallelism.