China Telecom's Xing4.0-29B-A4B: 29B MoE, 4B Active, 256K Context, Trained Fully on Ascend NPUs

China Telecom Artificial Intelligence Technology Co., Ltd. has released Xing4.0-29B-A4B, a mixture-of-experts language model with 29B total parameters and 4B activated per token, a native 256K context window (extensible to 512K), and training done entirely on Huawei Ascend NPUs. The company claims it is the first model of this scale trained fully on the Ascend platform with the MindSpore framework. This coverage is based on a GGUF quantization repository published by Venastine-Research; the base weights are listed under XingChen-AGI/Xing4.0-29B-A4B.

Xing is the successor branding to China Telecom's TeleChat series.

Architecture

According to the model card:

• Parameters: 29B total, 4B active per token
• Layers: 40, hidden size 3584
• Attention: Multi-head Latent Attention (MLA)
• Experts: 64 routed, 4 active per token, plus 1 shared; expert intermediate size 1024, dense FFN intermediate size 9216
• Context: 256K native, extensible to 512K
• Design: mHC + MLA + MTP (multi-token prediction), which the company says targets multi-step planning, tool calling and long-context coherence

Training ran on Ascend 910C clusters using MindSpore/MindFormers. China Telecom claims fine-grained MoE communication optimization, selective recomputation, DVM graph-operator fusion and Ascend C mHC fused operators raised training throughput by about 96% over out-of-the-box performance.

Benchmarks (self-reported)

The model card compares Xing4.0-29B-A4B with Gemma4-26B-A4B and Qwen3.6-35B-A3B. All figures come from China Telecom and have not been independently verified.

Benchmark	Xing4.0-29B-A4B	Gemma4-26B-A4B	Qwen3.6-35B-A3B
SWE-bench Verified	75.00	53.00	76.00
SWE-bench Multilingual	66.00	51.00	67.20
Terminal-Bench 2.1	57.50	30.00	51.50
Claw-Eval	76.55	71.49	74.54
Tau3-Bench	64.63	58.90	67.20
DeepresearchBII	60.80	39.30	59.70
AIME2026	90.00	88.30	92.70
IFBench	69.67	72.67	65.50
AA.LCR	61.00	66.00	62.00

The model leads Qwen3.6-35B-A3B on Terminal-Bench 2.1, Claw-Eval, DeepresearchBII and IFBench, and trails it on the other five. It trails Gemma4-26B-A4B only on IFBench and AA.LCR. The SWE-bench runs use the SWE-agent harness with a 210K context window. Several scores are averages over 3 to 5 runs, and Terminal-Bench 2.1 uses a 24-hour timeout. Harness differences can shift agentic scores materially, so treat cross-model comparisons with caution.

Deployment and quantizations

The weights are in Hugging Face Transformers format and work with vLLM, SGLang and KTransformers. Fine-tuning is supported through LLaMA-Factory and MindFormers. The card says it has format alignment for agent frameworks including OpenCode, Claude Code, OpenClaw and Hermes. A thinking mode is toggled with enable_thinking in the chat template. Recommended sampling is temperature 1.0 for general and reasoning tasks and 0.8 for coding and agent tasks, with top_p 0.95 and repetition_penalty 1.05.

The Venastine-Research GGUF repo lists file sizes by quantization:

• IQ2_M: 9.9 GB
• IQ3_XXS: 11.6 GB (MTP variant also 11.6 GB)
• IQ3_M: 13.9 GB (MTP variant 13.1 GB)
• Q4_K_M: 19 GB
• Q5_K_M: 22.2 GB
• Q6_K: 25.7 GB
• Q8_0: 33.2 GB
• F16: 62.5 GB

The repo showed 11,013 downloads in the last month. No Inference Provider hosts it, and API pricing is not yet disclosed. The license, training cutoff date and training token count are not stated in the available material.

What this means

The most notable claim is the training stack: a competitive agentic-coding model, by its own numbers, trained entirely on Ascend 910C hardware without NVIDIA GPUs. If the results hold up independently, it shows that domestic Chinese hardware can support frontier-adjacent MoE training, though the 96% throughput gain is measured against an unoptimized baseline, not against a GPU cluster.

For builders, the practical appeal is the 4B active footprint. The Q4_K_M file at 19 GB puts a 256K-context agentic model within reach of a single high-memory consumer GPU or a unified-memory machine. Its self-reported edge over Qwen3.6-35B-A3B is narrow, and it is behind on SWE-bench Verified and Tau3-Bench, so independent evaluation on matching harnesses should come before any switch.