ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

ZGCM-1 is a 7B dense foundation model utilizing interleaved gated sliding-window and full attention architectures to handle 256K context lengths. It uses a stable FP8 Muon optimizer and a progressive curriculum scaling from 16K to 256K context. The training recipe integrates Markov Decision Process reformulations of interaction traces to enhance agentic tool use.