China Telecom's XingChen lab trains 285B model on 8,192 Ascend 910C chips, open-sources 29B MoE
XingChen-AGI · hf · 2026-11-18
XingChen AGI Lab (China Telecom) released a technical report on Xing 4, a TeleChat-lineage model series trained end-to-end on Ascend NPUs with MindSpore.
- Architecture: MLA + DSA with cross-layer IndexShare for long context; wider-shallower MoE topology; block-wise draft model with lightweight correction heads for speculative decoding. A 29B MoE member (4B active params/token, native 256K context) is open-sourced.
- Pretraining infra: Built on MindSpore HyperParallel with topology-aware hybrid sharding, distributed optimizer execution, and CPU affinity/isolation, supporting a 285B model trained on 8,192 Ascend 910C cards.
- Post-training: TeleGym, a unified agent execution platform linking execution, evaluation, and trajectory construction to policy optimization, with disaggregated async execution, bounded oversampling, and partial rollout.
More from Infra
- Stratechery: Nvidia's earnings are a statement; Microsoft and Anthropic ship dueling AI OS updates — Stratechery · 2026-11-18
- Codeberg goes dark: domain vanished, pages unconfigured — jedisct1 · 2026-11-18
- All hyperscalers but Microsoft post negative FCF as AI capex peaks — marlene_zw · 2026-11-18
- Codeberg's domain mysteriously vanishes from its registrar account, site down — kingslyj · 2026-11-18
- Tokkey debuts thumb-sized AI hardware to run local models on your Mac — TejasKumar_ · 2026-11-18
- Hugging Face ships caching and serving upgrades, TTFT 1.5x faster — osanseviero · 2026-11-18