Zhongguancun Opens ZGCM-1 7B, an MIT Math and Agentic Search Model

Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence released ZGCM-1-7B on Hugging Face with a last-modified timestamp of September 16, 2026, shipping a 7.39B dense model under the MIT license alongside training code, staged checkpoints, and a public dataset card. The accompanying arXiv report (2609.13356, submitted September 11, 2026) frames the drop as a fully open recipe for math reasoning and tool-assisted search with 256K context.

ZGCM-1 is a decoder-only dense Transformer that interleaves gated sliding-window attention with global layers, and the Hugging Face card requires trust_remote_code=True because the project ships custom modeling code. The card and report describe thinking and direct-response modes in one checkpoint, plus multi-step tool use for web research and binary function search.

What shipped

Public artifacts include huggingface.co/zgcagi/ZGCM-1-7B (MIT-tagged weights), zgcagi/ZGCM-1-Data, and github.com/zgcagi/ZGCM-1 (MIT LICENSE verified). The model card reports author-measured scores such as 97.13% on MATH-500, 75.00% on AIME 2026, 63.09% on WebWalkerQA, and 62.00% on Binary Function Search; treat those figures as author-reported evaluations from the technical report and card, not independent Writeble benchmarks.

Primary sources are the Hugging Face model card, the arXiv abstract for 2609.13356, the MIT-licensed GitHub training repository, and the ZGCM-1-Data dataset card.