China Telecom's Xing4.0-29B-A4B lands llama.cpp support: trained fully on Ascend NPUs
jacek2023 · reddit · 2026-11-06
PR #31589 adds llama.cpp support for Xing4.0-29B-A4B, a 29B-total/4B-activated MoE model from China Telecom's AI unit (formerly the TeleChat series), with native 256K context extendable to 512K. It's billed as the first model of this scale trained entirely on Ascend 910C NPUs with MindSpore.
Key details:
- Architecture: mHC + MLA + MTP, agent-oriented with multi-step planning and tool calling
- Training efficiency: 96% throughput gain over out-of-the-box performance via MoE communication optimization, selective recomputation, DVM graph-operator fusion, and Ascend C fused kernels
- Ecosystem: fine-tuning via LLaMA-Factory/MindFormers; inference via SGLang/vLLM/KTransformers; adapted for OpenCode, Claude Code, and other agent frameworks
- Positioning: lightweight domain fine-tuning for intent classification, table understanding, contract auditing
A notable example of a frontier-scale open model trained end-to-end on a domestic Chinese compute stack.
More from Infra
- SpaceX data center pledges 40% power discounts and billions in community benefits — Scobleizer · 2026-11-06
- Mystery seller seeks to dump $500 million of data center loans — dinabass · 2026-11-06
- Collinear AI runs full-scale agentic RL on Together Post-Training with self-onboarding — togethercompute · 2026-11-06
- SemiAnalysis: Google TPUv7 Delivers Up to 50% Better Performance-per-Dollar Than B200/B300 for Inference — vllm_project · 2026-11-06
- Meta President: new US data center build-outs will be 'very challenging' — SumitGup · 2026-11-06
- Veteran investor: memory stocks aren't cheap even at P/E of 3 as LTAs are meaningless — bookwormengr · 2026-11-06