China Telecom's Xing4.0 trends on HF: 29B MoE trained entirely on Ascend 910C
AdinaYakup · x · 2026-09-26
Xing4.0 from China Telecom has been quietly trending on Hugging Face since last week. Key specs: a 29B/4B MoE model trained entirely on Huawei Ascend 910C with MindSpore (no NVIDIA GPUs), 256K context (up to 512K), released under Apache 2.0 — another concrete example of an open model trained on a domestic Chinese compute stack.
Related event: China Telecom Open-Sources Xing4.0, Trained Fully on Huawei Ascend Chips(2 posts)→
More from Infra
- Meta may expand Muse to 1B users, each agent sandbox costing ~$0.33/hour at market rates — sergeykarayev · 2026-09-26
- DuckDB queries Hugging Face datasets in place via the hf:// protocol — lhoestq · 2026-09-26
- First full 3D nuclear simulation didn't arrive until 2002, amid massive compute gaps — lauriewired · 2026-09-26
- OpGuard pinpoints bitwise divergence to debug LLM training at PyTorch Conference — PyTorch · 2026-09-26
- New local AI app hits 164 tokens/sec on iPad, tests to follow — TheOyinbooke · 2026-09-26
- How should LLM apps handle provider timeouts mid-stream? — Suresh-R · 2026-09-26