TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks
Xingchen AGI Lab reports pretraining a 285-billion-parameter model on 8,192 Ascend 910C cards.
The Xing 4 report describes a MindSpore HyperParallel training system that coordinates distributed computation, long sequences and activation memory. It studies how architecture choices balance model capacity against compute, memory access and execution overhead. For agent post-training, the team introduces TeleGym, connecting task execution, evaluation and trajectory construction to policy optimization. The supplied abstract gives no benchmark scores establishing frontier-level performance.
HF Daily Papers' note
The Xing 4 report describes a MindSpore HyperParallel training system that coordinates distributed computation, long sequences and activation memory. It studies how architecture choices balance model capacity against compute, memory access and execution overhead. For agent post-training, the team introduces TeleGym, connecting task execution, evaluation and trajectory construction to policy optimization. The supplied abstract gives no benchmark scores establishing frontier-level performance.
HF Daily Papers' note
score 5