ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search
| Source: arXiv AI
Tags: ZGCM-1, open-source model, foundation model, agentic search, long-context, math reasoning
ZGCM-1 is a fully open 7B foundation model that releases weights at all training stages, training code, intermediate checkpoints, per-stage data, and W&B logs—achieving competitive performance with Qwen3-235B on math and agentic search while demonstrating a ~4.2x training efficiency improvement at 16K context.
Details
Open-source foundation models tend to release weights without training details, making genuine replication difficult. ZGCM-1 from a 22-author team takes the opposite approach: everything is released. The 7B dense model features interleaved gated sliding-window and full attention enabling efficient 256K context, a stable FP8 Muon optimizer, and progressive curriculum training across 16K, 64K, and 256K context windows. Mid-training reformulates interaction traces as Markov Decision Processes, which the authors argue is key to competitive agentic search performance. On math and agentic benchmarks, ZGCM-1-7B remains competitive with frontier models orders of magnitude larger—including Qwen3-235B-A22B and GLM-5.1—while the 16K pre-training time-to-loss is ~4.2x more efficient than comparable approaches. The full release includes model weights from pre-training, mid-training, and post-training stages, intermediate checkpoints, training code, per-stage data and recipes, and W&B training logs. The paper distills eight empirical findings covering architectural scaling, SFT quality pruning, long-context generalization, and agentic co-training dynamics. An AI-native R&D workflow using agent swarms for cluster operations and data curation is also documented. This is one of the most complete open releases in the 7B model class.