Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence released ZGCM-1, a small AI model, on Sept. 11, 2026. It has 7.39 billion parameters, the numbers a model tunes as it learns. It was trained from scratch for mathematical reasoning and for agentic search, meaning it can carry out multistep searches on its own with outside tools. Because the team released the model's files from every stage of training along with its data, code and logs, other researchers can inspect how it was built and train it again, instead of only using a finished chatbot.
The authors say ZGCM-1 works through problems step by step and calls on outside tools as it goes, rather than relying only on what it picked up from web text during training. It was trained in stages to handle longer and longer material at once. The authors also report that their training design made the first, broad stage of training about 4.2 times more efficient.
On several hard math and search tests, the authors say, ZGCM-1 holds its own against far larger models such as Qwen3-235B-A22B and GLM-5.1, though on selected tasks rather than across the board. The model's page on Hugging Face lists high scores on math-competition tests and on a multistep web-search test.
The model files are on Hugging Face under the MIT license, a license that lets anyone use, change and share them free of charge. The training data is on Hugging Face too, and the training code is on GitHub.
So far, every score comes from the team itself. No outside group has tested the model yet, so it is too early to say how well a model this small really stands up to much larger ones.