Xing4.0-29B-A4B: China Telecom Lab's Open-Weight MoE Trained on Ascend Chips
Xing4.0-29B-A4B is an Apache 2.0 mixture-of-experts model from China Telecom's Xingchen AGI Lab, with 29B total and 4B active parameters per token and a native 256K context. A technical report posted on 17 November 2026 says its 285B sibling was trained on 8,192 Huawei Ascend 910C chips. GGUF builds of the open model range from 11 to 34 GB.
On this page
What the Xing4 report announced
On 17 November 2026, a team listed as the Xingchen AGI Lab, with 138 authors, posted a technical report to arXiv titled "TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks". The Hugging Face model card names the developer as China Telecom Artificial Intelligence Technology and says the series was formerly called TeleChat.
The report covers two models. The flagship is a 285B-parameter mixture-of-experts model with about 17B parameters active per token, which has not been released. The open member is Xing4.0-29B-A4B, which activates only 4B of its 29B parameters per token and supports a native 256K-token context. The weights are published on Hugging Face under the Apache 2.0 license, with BF16, FP8, W8A8 and GGUF versions.
Trained entirely on Huawei chips
The headline claim is about hardware. According to the abstract, the 285B model was pretrained on 8,192 Huawei Ascend 910C cards using MindSpore HyperParallel, a Chinese training stack. Much of the report reads as an engineering log of making that work: a wider, shallower layer layout chosen because it trains about 10% faster on Ascend clusters, changes to parallelism, and memory work to cut step times.
This matters beyond China Telecom. China Telecom already trained earlier TeleChat mixture-of-experts models entirely on Huawei chips, and US export controls keep pushing Chinese labs toward domestic chips. A competitive open model produced on that stack shows the gap is narrowing.
Speed claims, and what they cover
The report makes large throughput claims, but read them carefully. It cites up to 4.68 times higher prefill throughput from its attention design compared with a plain MLA baseline, and 2.67 to 3.77 times faster decoding from a speculative decoding method called DSpark. The prefill figure comes from a simulator that the authors themselves say may be more optimistic than real deployment. The DSpark figures were measured on the unreleased 285B model, not on the 29B open model, and the open repository does not include a DSpark draft model.
On quality, the model card compares Xing4.0-29B-A4B with Gemma4-26B-A4B and Qwen3.6-35B-A3B. It reports 75.0 on SWE-bench Verified against 53.0 and 76.0, and 57.5 on Terminal-Bench 2.1 against 30.0 and 51.5. These are the developer's own numbers and have not been independently checked.
Running Xing4.0-29B-A4B locally
With only 4B active parameters, this is a model aimed at ordinary hardware. The full BF16 repository is about 62 GB, but the model card lists GGUF builds from 2-bit to 8-bit at 11 to 34 GB, with support merged into mainline llama.cpp. The card's collection notes that a 4-bit build needs about 15 GB of VRAM, within reach of a 24 GB RTX 3090 or 4090, and community MLX conversions exist for Apple silicon. SGLang support is merged and an official vLLM plugin is available.
Two practical notes. The Transformers path needs trust_remote_code, so review the custom modeling code before running it. And thinking mode is on by default, which lengthens responses. For local agent and coding work, a 4B-active model with a 256K context is a useful size, and Apache 2.0 allows commercial use.