MODEL

Xing4

modeltopic-notechinaopen-source

Overview

Xing4 is a family of LLMs evolved from TeleChat and designed for Huawei Ascend hardware, introduced in the arXiv:2611.22115 paper “TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks”. The headline artifact is a 285B-parameter model pretrained on 8,192 Ascend 910C cards, plus a TeleGym agent-execution platform for post-training. Fits the now-established Ascend training track that already includes Zhipu AI‘s GLM family, Huawei’s openPangu-2.0-Pro (505B, Aug), and the broader GLM-5 / GLM-5.3 arc the corpus has been carrying — a scale-continuity datapoint on the Chinese-domestic-silicon training stack, not a first-of-kind result.

Timeline

  • 2026-11-18-AI-Digest — Xing4 lands as a ▲85-upvoted HuggingFace paper (arXiv:2611.22115): “TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks”. The paper introduces Xing4 — a family of LLMs evolved from TeleChat — and designs it for Ascend hardware, including a 285B-parameter model pretrained on 8,192 Ascend 910C cards plus a TeleGym agent-execution platform for post-training. Why it matters: a full frontier-scale training + agent post-training stack running end-to-end on non-NVIDIA silicon — adds a trillion-scale datapoint to a now-established Ascend training track that already includes Zhipu AI‘s GLM family, Huawei’s openPangu-2.0-Pro (505B, Aug), and the broader GLM-5 / GLM-5.3 arc the corpus has been carrying. Softened framing: the “hardware-sovereignty push” narrative is now ~12-month-old, not a Nov surprise — the right read is scale continuity on the Ascend track, not a first-of-kind result. Log against MOC - Open Source Models and MOC - AI Infrastructure.

Key Developments

  1. 285B-Parameter TeleChat-Lineage Model Pretrained on 8,192 Ascend 910C Cards With TeleGym Post-Training Platform (November 18, 2026): The arXiv paper (▲85 on HuggingFace, arXiv:2611.22115) documents a 285B-parameter Xing4 model pretrained end-to-end on 8,192 Ascend 910C cards and a TeleGym agent-execution platform for post-training. The load-bearing structural fact: a frontier-scale training + agent post-training stack running on non-NVIDIA silicon, extending the already-established Ascend training track (Zhipu AI GLM family, Huawei openPangu-2.0-Pro 505B Aug, GLM-5 / GLM-5.3). The right read is scale continuity on the Ascend track, not a first-of-kind hardware-sovereignty result — the “domestic-silicon training stack” narrative is now ~12-month-old, not a November surprise.

See also: Zhipu AI, Z.ai, GLM 5.3, GLM 5.2, NVIDIA, MOC - Open Source Models, MOC - AI Infrastructure.