> ## Content Index
> Fetch the complete content index at: https://www.notatechguy.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Xing 4: China's 285B AI model on 8,192 Ascend cards
- URL: https://www.notatechguy.com/xing-4-china-s-285b-ai-model-on-8-192-ascend-cards/
- Published: 2026-11-18T14:08:53.000Z
- Updated: 2026-11-18T14:08:53.000Z
- Description: Xingchen AGI Lab trained a 285B model on 8,192 Huawei Ascend cards but published no benchmark scores or comparisons against Western models.
- Author: Marcello Babbili
- Tags: Technology & AI, OpenAI, AI Agents, Google

Xingchen AGI Lab trained a 285-billion-parameter language model across 8,192 Ascend 910C cards, according to a paper posted November 18 [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). The paper includes no benchmark scores and no comparisons against Western models, and it has not been peer-reviewed [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). Ascend is Huawei's domestic chip line and MindSpore is Huawei's AI framework, making the work a test of China's push to build large models on its own hardware.

**My read:** I don't buy the "frontier" label in the paper's title. The infrastructure is impressive: 8,192 Ascend cards running a 285B model is serious scale. But without a single benchmark score or comparison to GPT-4, Claude, or even open models like Llama, "frontier" is a self-description, not a demonstrated fact. The number I would watch is whether any third party downloads the released 29B variant and publishes scores. Until then, this is a systems paper, not a capabilities paper.

### What was actually built

The Xing 4 series extends the earlier TeleChat models [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). The paper, posted to arXiv by the Xingchen AGI Lab team including Baotian Hu and Dongfang Li [P²](https://huggingface.co/papers/2611.22115?ref=notatechguy.com), covers architecture selection, large-scale pretraining, and post-training for reasoning and agentic tasks [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com).

On architecture, the team studied how sparse activation, attention design, model depth and width, and speculative decoding trade off between model capacity and the real costs of computation, memory access, and execution overhead [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). Sparse activation is the technique behind mixture-of-experts models, where only a fraction of the network fires for any given token, cutting the cost of running very large models. How those weights are structured and modified matters, as we found when [AI models scored 53% editing neural network weights directly](https://www.notatechguy.com/ai-models-score-53-editing-neural-network-weights-directly/).

The training infrastructure runs on MindSpore HyperParallel, Huawei's distributed training framework [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). It combines topology-aware hybrid sharding, which splits the model across chips based on how they are physically connected, with efficient distributed optimizer execution and coordinated management of long-sequence workloads and activation memory [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). The team also used CPU affinity and core isolation to reduce scheduling interference on the host machines [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com).

The headline scale is a 285B-parameter model pretrained on 8,192 Ascend 910C cards, a run nobody outside the lab has verified [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com).

### A smaller model you can download

While the paper describes the 285B training run, a specific variant called Xing4.0-29B-A4B is available on Hugging Face [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com). It is a mixture-of-experts model with 29 billion total parameters and 4 billion active parameters, with a native 256,000-token context window [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com). The model page describes it as shifting from "answering questions" to "getting things done," positioning it as an agentic model rather than a chatbot [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com).

The release includes multiple build formats: BF16, FP8, and GGUF quantised versions ranging from 11 to 34 GB that can run in llama.cpp [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com). There is also a vLLM plugin and deployment options on Tianyi AI Platform and ModelScope [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com).

No one outside the lab has independently benchmarked this model.

### TeleGym: training agents, beyond chatbots

The paper introduces TeleGym, a unified agent execution platform that links agent execution and evaluation to trajectory construction and policy optimization [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). TeleGym uses disaggregated asynchronous execution, bounded oversampling, and partial rollout to handle tasks that take wildly different amounts of time while keeping trajectory continuity intact as the underlying policy updates [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com).

In practice, this means the system can train an agent on a mix of fast and slow tasks without losing the thread when the model itself changes mid-training. Agent training is where the field is heading, as we found when [AI models were shown to bend advice to serve the boss's bonus](https://www.notatechguy.com/ai-models-bend-advice-to-serve-the-boss-s-bonus/), a reminder that agentic behaviour introduces reliability questions that pure text generation does not.

### The broader picture

Xing 4 is not an isolated project. Baidu's Qianfan-VL, a multimodal model also trained on domestic chips, was described in a September 2025 report as achieving state-of-the-art performance on domain-specific tasks [P⁵](https://huggingface.co/blog/baidu/qianfan-vl-introduction?ref=notatechguy.com). The pattern is consistent: Chinese AI labs building on Huawei Ascend silicon and MindSpore frameworks to reduce dependence on NVIDIA hardware and Western software stacks.

This comes as global AI infrastructure continues to shift, with [OpenAI bringing frontier models to Google Cloud](https://www.notatechguy.com/openai-brings-frontier-models-to-google-cloud/) in October, a move that showed even the largest Western labs are diversifying their compute.

### What is missing

The paper reports no evaluation scores and no comparisons against other models [S¹](https://arxiv.org/abs/2611.22115?ref=notatechguy.com). The 285B parameter count and the 8,192-card training scale are self-reported. The paper has not been peer-reviewed. Whether the released 29B variant performs competitively against similarly sized open models remains an open question that only independent benchmarking can answer.

For a team evaluating whether to deploy Xing4.0-29B-A4B in a Chinese-market application, the practical first step is to download the GGUF build from Hugging Face, run it through a standard agent benchmark in llama.cpp, and compare against an existing model of similar size. The Xing4.0-29B-A4B weights are available now on Hugging Face in BF16, FP8, and GGUF formats, with the smallest quantised build fitting in 11 GB of memory [P³](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com).

---

*Sources: [S1 — TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips](https://arxiv.org/abs/2611.22115?ref=notatechguy.com) · [P2 — Paper page - TeleChat to Xing4: Advancing Frontier Agentic Models on D](https://huggingface.co/papers/2611.22115?ref=notatechguy.com) · [P3 — XingChen-AGI/Xing4.0-29B-A4B · Hugging Face](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B?ref=notatechguy.com) · [P4 — AgentFrontier: Expanding the Capability Frontier of LLM Agents with ZP](https://arxiv.org/pdf/2510.24695v1.pdf?ref=notatechguy.com) · [P5 — Qianfan-VL: A Milestone Achievement in Chinese Multimodal AI with Dome](https://huggingface.co/blog/baidu/qianfan-vl-introduction?ref=notatechguy.com)*

## Related reading

- [OpenAI brings frontier models to Google Cloud](https://www.notatechguy.com/openai-brings-frontier-models-to-google-cloud/) — our technology desk, 2026-10-21
- [AI models bend advice to serve the boss's bonus](https://www.notatechguy.com/ai-models-bend-advice-to-serve-the-boss-s-bonus/) — our technology desk, 2026-11-13
- [AI models score 53% editing neural network weights directly](https://www.notatechguy.com/ai-models-score-53-editing-neural-network-weights-directly/) — our technology desk, 2026-11-16

---

*Written from 5 sourced items, 4 of them primary.*

![Xing 4 model sizes: pretrained vs released](https://storage.ghost.io/c/6e/89/6e896869-22ef-4281-a213-b4c462c17cff/content/images/2026/11/chart_3fe93bff7540251f804b.png)