China Telecom open-sources Xing4.0-29B MoE, first in class trained fully on Ascend NPUs
Skyline34rGt · reddit · 2026-09-17
China Telecom AI has released Xing4.0-29B-A4B on Hugging Face, the next-gen model of its Xing series (formerly TeleChat): a 29B-total-parameter MoE with only 4B active per token.
Key points
- Architecture: mHC + MLA + MTP, native 256K context (extensible to 512K), 64 routed experts + 1 shared expert, designed for multi-step planning, tool calling, and agent workloads.
- Domestic training: claimed first model at this scale trained entirely on Ascend 910C clusters with MindSpore; co-optimizations (MoE communication, selective recomputation, DVM graph-operator fusion, Ascend C fused kernels) boosted training throughput 96% over out-of-the-box.
- Ecosystem: fine-tuning via LLaMA-Factory/MindFormers, inference via SGLang/vLLM/KTransformers, and adaptation for agent frameworks like OpenCode and Claude Code.
- Benchmarks: SWE-bench Verified 75.0, Terminal-Bench 2.1 57.5 (beating compared Gemma4-26B-A4B), AIME2026 90.0, roughly on par with Qwen3.6-35B-A3B overall.
More from Infra
- To dethrone Postgres, new databases must prioritize ecosystem integrations from day one — mgill25 · 2026-11-02
- ESP32-powered 4-inch touchscreen builds a TUI dashboard for monitoring a DGX inference cluster — CATLLM · 2026-11-02
- OpenRouter crosses 10 quadrillion annualized token run rate, doubling every 10.5 weeks — deedydas · 2026-11-02
- Palmer Luckey asks: can you use datacenter RAM for things prohibited by the operator's terms of service? — avaitopiper · 2026-11-02
- Disable MTP for 20-30% faster local LLM inference on limited RAM — InternalMode8159 · 2026-11-02
- Kioxia up 216% in a year but down 68% from its June 2026 peak — zephyr_z9 · 2026-11-02