The paper presents ZGCM-1, a fully open 7B dense foundation model trained from scratch to couple internal reasoning with external tool use rather than relying on parametric memorization. Its training recipe combines interleaved gated sliding-window/full attention, an FP8 Muon optimizer, and progressive context scaling to 256K tokens with interaction traces reformulated as Markov Decision Processes, yielding a 4.2x efficiency gain in pretraining time-to-loss. ZGCM-1-7B is competitive with frontier models tens of times larger (e.g., Qwen3-235B-A22B, GLM-5.1) on mathematical reasoning and agentic search benchmarks, and the authors open-source weights, code, data recipes, and checkpoints across all training stages.
