ZGCM-1: An Open, Efficient 7B Model for Math and Agentic Search

Summary

The paper introduces ZGCM-1, a fully open 7B dense foundation model trained from scratch around the idea that a compact model can compensate for limited parametric capacity through deliberate internal reasoning and active external tool use. The model supports a 256K context window. Its training recipe combines interleaved gated sliding-window and full attention with a stable FP8 Muon optimizer, progressive context scaling from 16K to 64K and 256K, and mid-training that reformulates interaction traces as Markov Decision Processes. The authors also describe an AI-native research workflow in which agent swarms manage cluster operations, curate data, and run diagnostic evaluations. In reported evaluations, ZGCM-1-7B is competitive with other 7B models on general benchmarks and remains competitive with much larger frontier models, including Qwen3-235B-A22B and GLM-5.1, on selected mathematical-reasoning and agentic-search suites. The pre-training design produced an approximately 4.2-fold improvement in 16K pre-training time-to-loss. The paper distills eight empirical findings covering architectural scaling, supervised fine-tuning data pruning, long-context generalization, and agentic co-training dynamics. To support reproducibility and community research, the authors release weights from pre-training, mid-training, and post-training, intermediate checkpoints, training code, per-stage data and recipes, and Weights & Biases logs.