ZGCM-1-7B: Architecture, Training, Benchmarks, and License

ZGCM-1-7B is a dense language model trained from scratch to combine mathematical reasoning with active, tool-assisted information gathering. Researchers at Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence released it as a model that others can inspect and test, rather than only compare through benchmark scores. 24

Release and license

The model has 7.39 billion parameters. Its model listing identifies the release as MIT-licensed; the CC BY 4.0 label on the research paper applies to the paper and should not be mistaken for the model’s license. Researchers should check the terms attached to each resource they use. 271

How the model works

ZGCM-1-7B is a dense, decoder-only Transformer with a stated 256K-token context window. It supports both a deliberate thinking mode and a direct-response mode in the same model, making it possible to study different response styles without changing models. 24

Its attention design interleaves gated sliding-window layers with full-attention layers; one model description specifies 27 of the former and five of the latter. The authors report that the hybrid design cuts KV-cache use by 6.4× and increases throughput at 256K context by 3.94× relative to their full-attention comparison. Those are comparison-specific system results, not a promise of the same speedup on every deployment. 410

What makes the training release useful

The published recipe runs from pretraining through progressive long-context mid-training and general-agentic supervised fine-tuning. It combines hybrid attention with a stable FP8 Muon optimizer, extends the curriculum across 16K, 64K, and 256K stages, and reformulates interaction traces as Markov decision process transitions. The reported mid-training curriculum covers 600 billion tokens. 1210

Alongside the final model, the project has published a dataset and intermediate checkpoints; reporting on the release also describes training code, recipes, and logs. That gives researchers more of the training path to examine than final weights alone. A checkpoint labeled for the 16K data stage, for example, documents its training context and cumulative mid-training tokens. 2369

The team describes researcher-directed AI agents assisting with development tasks, including data curation and cluster operations. That is a description of its workflow—not evidence that agents independently designed or validated the model, or a measurement of how much they contributed to its scores. 28

Reported results—and their limits

The reported evaluations give ZGCM-1-7B 97.13% on MATH-500, 75.00% on AIME 2026, 63.09% on WebWalkerQA, and 19.43% on BrowseComp. Results are not uniformly ahead of comparison models: the published comparisons also show losses on AIME 2024 and AIME 2025. Separately, the authors report reaching the same pretraining loss in roughly 4.2× less time than their AdamW/BF16 baseline. These figures depend on the stated benchmarks, evaluation setup, and system baselines; they do not establish equivalent performance in every search workflow. 710

Getting started

The model listing provides a Transformers text-generation example for zgcagi/ZGCM-1-7B that uses trust_remote_code=True. Review the repository’s custom code before enabling that option. The provided material does not establish a minimum GPU, memory requirement, or package-version requirement for running the final model, so those should not be inferred from the benchmark or checkpoint descriptions. 219