Our Blog

Blog Index 

China Telecom's Xing4 Report: A 285B-Parameter Model Trained on 8,192 Huawei Ascend Chips

Posted on 18th Nov 2026 11:03:38 in Artificial Intelligence, Machine Learning

Tagged as: China Telecom, Huawei Ascend, Xing4, AI chips, MindSpore, open weights

China Telecom's AI research arm has published a technical report describing one of the largest AI training runs carried out entirely on Chinese silicon: a 285-billion-parameter model pretrained on 8,192 Huawei Ascend 910C accelerators, using the company's own MindSpore HyperParallel framework rather than Nvidia GPUs and CUDA.

The report, titled TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks (arXiv:2611.22115), was submitted on November 17, 2026, and was ranked as Hugging Face's number one paper of the day on November 18. It was led by Xingchen AGI Lab at China Telecom, in collaboration with Huawei and the Harbin Institute of Technology (Shenzhen), and lists 138 authors.

The work is not a consumer chatbot launch. It is a systems paper: how a Chinese telecom operator staged a frontier-scale MoE pretraining run on domestic accelerators, how it tuned the software stack around them, and how it built agent post-training infrastructure on top. That makes it a useful marker of how far the non-Nvidia AI stack has actually progressed, and where it still depends on unverified claims.

What China Telecom Actually Built

Xing4 is a family of models continuing the earlier TeleChat lineage, spanning tens to hundreds of billions of parameters with stated ambitions toward trillion-parameter scale. The report documents two representative configurations.

The first is Xing4.0-29B-A4B, the open-weight member. It carries 29 billion total parameters but activates only about 4 billion per token, roughly 13.8% activation. It uses 40 layers, a hidden size of 3,584, a dense feed-forward width of 9,216 and an expert intermediate size of 1,024. Routing sends each token through four of 64 routed experts plus one shared expert. Attention is Multi-head Latent Attention (MLA), combined with multi-token prediction and Manifold-Constrained Hyper-Connections (mHC). The context window is 256K tokens natively, extensible to 512K, and it was trained on 1.5 trillion bilingual Chinese and English tokens. The weights are released under Apache 2.0.

The second is the 285B-parameter flagship, the model behind the headline number. Its configuration differs: a wider-and-shallower MoE topology, MLA paired with DeepSeek Sparse Attention (DSA) and cross-layer IndexShare for long-context serving, optimised mHC operators, and a block-wise draft model with lightweight correction heads for speculative decoding, a technique the paper calls DSpark.

Post-training is handled by TeleGym, a unified agent execution platform that connects environment execution, evaluation, trajectory construction and policy optimisation. Three design choices stand out: disaggregated asynchronous execution, so fast tasks do not wait on slow tools; bounded oversampling to control cost on long-tailed task durations; and partial rollout, which preserves trajectory fragments across policy updates instead of discarding half-finished episodes.

Inside the Ascend Training Stack

The paper is candid about why this is hard. Its opening line notes that most open-source models are trained on GPUs, and that scaling on domestic chips and frameworks poses significant systems challenges. The answer is MindSpore HyperParallel, which combines topology-aware hybrid sharding, efficient distributed optimizer execution, and coordinated management of long-sequence workloads and activation memory. The authors also highlight CPU affinity and core isolation, a host-level tuning step that reduces scheduler jitter, which can otherwise desynchronise collective operations at cluster scale.

The systems results are specific and therefore checkable, at least in principle:

  • Up to 4.68x prefill throughput over an MLA-only baseline, credited to the sparse attention and index-sharing design.
  • 2.67x to 3.77x decoding throughput from DSpark speculative decoding.
  • 16.1% throughput improvement from eliminating pipeline parallelism.
  • Roughly 96% training throughput gain over out-of-the-box performance, from fine-grained MoE communication optimisation, selective recomputation, automatic graph-operator fusion and hand-tuned fused mHC kernels.

All of these are author-reported measurements made on China Telecom's own infrastructure. None has been independently reproduced on third-party Ascend clusters, and the paper does not publish an MLPerf-style apples-to-apples comparison against Nvidia hardware.

Benchmarks: Where the Open 29B Model Lands

The 29B model card provides head-to-head numbers against two same-class open agentic models, Gemma4-26B-A4B and Qwen3.6-35B-A3B. On coding-agent work, the results are competitive rather than dominant:

  • SWE-bench Verified: 75.0 for Xing4.0, against 53.0 for Gemma4-26B-A4B and 76.0 for Qwen3.6-35B-A3B.
  • Terminal-Bench 2.1: 57.5 for Xing4.0, against 51.5 for Qwen3.6 and 30.0 for Gemma4.
  • SWE-bench Multilingual: 66.0, against 51.0 and 67.2 respectively.
  • Agent benchmarks: Claw-Eval 76.55 and DeepresearchBII 60.80, both the best in that three-model comparison; Tau3-Bench 64.63, behind Qwen3.6's 67.20.
  • Reasoning: AIME2026 at 90.00, behind Qwen3.6's 92.70; IFBench 69.67, behind Gemma4's 72.67.

The pattern is a model that is strong on terminal and long-horizon agent tasks, roughly level on software engineering, and slightly behind on some reasoning and instruction-following suites. China Telecom evaluated SWE-bench Verified with the SWE-agent harness at a 210K context window, with temperature 1.0 and top-p 0.95.

Running It Without Huawei Hardware

This is where the release becomes practically interesting for developers outside China. Although the training stack is Ascend-specific, inference has been ported into mainstream open-source serving software.

According to the model card, SGLang added native support in mainline on November 5, 2026. llama.cpp merged native GGUF support in the same window, usable from stock builds b12174 onward on CPU or CUDA. vLLM support ships as an official out-of-tree plugin, while TensorRT-LLM and KTransformers integrations are still pull requests under upstream review. Quantised builds range from 2-bit to 8-bit GGUF at roughly 11 to 34 GB, and a 4-bit quantisation is quoted at around 15 GB of memory for long-context local runs.

The model is also format-aligned with agent harnesses including OpenCode, Claude Code, OpenClaw and Hermes, and fine-tuning recipes have been merged into ms-swift and LLaMA-Factory. In plain terms: an Apache 2.0 agentic coding model with a 256K context window, trainable and servable on ordinary GPUs, is downloadable now. That is a different proposition from a paper-only claim.

Why the Timing Matters

The report landed in the same week that Nvidia confirmed it has stopped shipping its H200 and other Hopper-series AI chips to China. Interim CEO Colette Kress disclosed the halt on the Q3 earnings call, ending an arrangement brokered by former CEO Jensen Huang and President Donald Trump that routed sales through the United States and collected a 25% tariff. Reuters reported earlier in November that around 600,000 H200 chips were sitting in Hong Kong awaiting shipment to the mainland while Beijing withheld approval for large-scale imports. Washington still bans exports of Nvidia's more advanced Blackwell and Rubin GPUs to China.

Against that backdrop, a 285B-parameter training run on domestic accelerators is a strategic data point, not just an engineering one. China Telecom has been building toward this for years: in January 2026, the South China Morning Post reported that its TeleChat3 MoE models were the country's first trained entirely on Huawei's Ascend 910B chips. Xing4 moves the same approach up a generation, to 910C silicon and to a much larger parameter count.

For Indian teams, the read-through is narrower but real. First, the model itself is usable: Apache 2.0 weights that run on llama.cpp and vLLM remove any hardware dependency for inference. Second, the benchmark table gives a concrete sense of where a domestically trained Chinese model sits relative to Google's and Alibaba's open releases, which matters when evaluating cheap coding agents for internal tooling. Third, the export-control squeeze that produced this stack is the same one reshaping global accelerator supply, and it is the reason Indian buyers face volatile pricing and allocation on high-end GPUs.

What Remains Unproven

Several claims deserve caution. The 285B model has not been released as weights; only the 29B member is open. The throughput multipliers, including the 96% training gain, are measured by the authors on their own cluster and have no independent replication. The benchmark comparisons are self-reported on the model card, with harness details published but no third-party audit. And the report does not answer the question most buyers actually care about: total cost per token at scale versus an Nvidia-based cluster, including power, cooling and utilisation.

There is also an ecosystem cost. Training on MindSpore and Ascend means Chinese operators can keep building without CUDA licences or US-approved hardware, but it also means checkpoints and kernels are tuned for a stack that most of the world's tooling does not target. Inference ports soften that, yet the training recipes remain Ascend-oriented.

What the paper does establish is narrower and more durable: frontier-scale pretraining on non-Nvidia silicon is no longer theoretical. A state-owned telecom operator, working with Huawei and a university lab, has documented a 285B-parameter run on 8,192 domestic accelerators, published the systems engineering behind it, and shipped a usable open-weight model alongside it. Whether that closes the gap with the American frontier is an evaluation question the report deliberately leaves open.

Sources

whatsapp me