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

ZGCM-1. A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search. By Jiyan He and twenty-one others, from Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence. Published on the eleventh of September. Its code repository is worked out: about three thousand five hundred source files, with a readme and a licence.

ZGCM-1 is a fully open dense foundation model with seven billion parameters, trained from scratch and aimed at math and agentic search. Fully open here means a long list of released artifacts. There are model weights from the pre- training stage, the mid-training stage and the post-training stage. There are intermediate checkpoints, the training code, the data for each stage together with its data recipe, and the W&B logs from training.

The model supports a context of two hundred fifty-six thousand tokens. It is built on one premise: a compact model cannot passively memorize the open web, but it can overcome the limits of its parametric capacity by coupling deliberate internal thinking with active external tool use. Rather than storing everything, it reasons and it looks things up.

On evaluation, ZGCM-1-7B is competitive within the family of seven billion parameter models on general benchmarks. On several challenging mathematical reasoning and agentic search suites, it remains competitive with frontier models orders of magnitude larger, with Qwen3-235B-A22B and GLM-5.1 named as examples.

The training recipe is built for efficiency in data, in systems and in algorithms, end to end. On architecture and systems, which were co-designed, it interleaves gated sliding-window attention with full attention, and it uses a stable FP8 Muon optimizer. Mid-training follows a progressive curriculum. Context is scaled in steps: sixteen thousand tokens, then sixty-four thousand, then two hundred fifty-six thousand. And interaction traces are reformulated as Markov Decision Processes.

The development itself ran on what the authors call an AI-native research and development workflow, in which agent swarms autonomously manage cluster operations, data curation, and rapid diagnostic evaluation. The pre-training design is reported to give about a four point two times efficiency improvement in pre-training time-to-loss at a sixteen thousand token context. Across the full development lifecycle, the authors distill eight actionable empirical findings, spanning architectural scaling, SFT quality pruning, long-context generalization, and the dynamics of agentic co-training.

The stated purpose of the release is community research. Because weights come from every stage, along with intermediate checkpoints, per-stage data and logs, the release serves someone studying how a compact model is trained for long context and tool use, not only someone who wants the finished model.

ZGCM-1 rests on a bet that a small model should think and reach for tools rather than try to hold the web in its weights. The release makes that bet inspectable from the first pre-training stage to the last post-training checkpoint.