Wednesday, 18 November 2026 · 6 min read
Xing 4 advances large language models on Ascend chips with efficient architecture and training
Xing 4 models optimize architecture and training on Ascend NPUs, achieving efficient long-context reasoning and agentic tasks with a 285B-parameter model pretrained on 8,192 Ascend 910C cards and a 29B open-source MoE model supporting 256K-token contexts.
Based on the paper TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks by Xingchen AGI Lab, Baotian Hu, Bingkai Yang, Boyang Wang and 135 others
Challenges of scaling large language models on domestic Ascend hardware
Scaling large language models (LLMs) to hundreds of billions of parameters requires balancing model capacity, long-context reasoning, and efficient execution, especially on domestic Ascend NPUs with distinct hardware constraints.
Long-context reasoning demands handling very large key-value caches for attention, which increases memory and latency costs, while Ascend hardware favors matrix-dominant computation but has limited vector compute and memory bandwidth.
Training at scale on thousands of Ascend NPUs introduces systems challenges such as communication overhead, memory pressure, and host scheduling interference, complicating efficient distributed training and inference.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- The key formula from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short