IDEA Prune: An Integrated Enlarge-and-Prune Pipeline in Generative Language Model Pretraining

Apple Research Blog
Read full post
Researchers propose IDEA Prune, an integrated pipeline combining enlarged model pretraining, pruning, and recovery under a unified learning rate schedule to improve token efficiency and performance in compressing large language models. Experiments show compressing 2.8B to 1.3B parameter models with up to 2T tokens benefits from this approach.

More on this story


More in LLM & Text Generation

Build more natural voice experiences with GPT‑Live‑1 in the API

Covered by 2 sources

DeepSeek V4.1 Flash now available on AI Gateway

Covered by 2 sources