Zhongguancun Academy opens ZGCM-1-7B weights, data and code

Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence have released ZGCM-1-7B, a 7.39-billion-parameter dense model built for mathematical reasoning and tool-assisted agentic search. Weights, training data and training code ship together under an MIT license on Hugging Face and GitHub, with a technical report documenting the from-scratch recipe.

The architecture is a hybrid attention layout of 27 gated sliding-window layers plus five global layers, paired with a 256K-token context window. A single checkpoint carries both a thinking mode and a direct-response mode, so the same weights serve long chain-of-thought reasoning and quick answers without separate deployments.

Pretraining consumed roughly 4.19 trillion tokens using hybrid FP8 precision and the Muon optimizer. Mid-training then extended context from 16K to 64K and finally to 256K over about 600 billion tokens, reformulating interaction traces as MDP state-action transitions, a step that treats agent tool use as a learnable sequential decision problem rather than a formatting task.

Supervised fine-tuning mixed general and agentic trajectories, with execution-verified tool use and assistant-only loss. Reported thinking-mode scores include about 97.13% on MATH-500, 75.00% on AIME 2026 and 70.42% on HMMT 2025, giving the model the best average rank among seven compared 7B-8B models across fourteen reasoning benchmarks in the report.

On agentic search the model reaches about 63.09% on WebWalkerQA with web search and page reading, 19.43% on BrowseComp, and 62.00% on Binary Function Search using Ghidra tools. BrowseComp in particular remains hard, showing how far a compact open model still sits from reliable multi-hop web research.

The report also makes architecture claims: roughly 3.94x training throughput at 256K context versus full attention, and about a 4.2x improvement in 16K pretraining time-to-loss from combined architecture, precision, optimizer and normalization choices. Those efficiency numbers matter more than leaderboard placement for labs that cannot afford full-attention long-context runs.

Operationally, the quickstart requires trust_remote_code to load custom modeling files, and the team publishes stage-specific configs for data processing, pretraining, mid-training, SFT and RL. The report also says researcher-directed AI agents assisted with data curation and evaluation.

For open-weight labs, ZGCM-1 matters less as a chat model than as an inspectable end-to-end pipeline for math reasoning and agentic search. The next things to watch are independent reproductions of the benchmark numbers and whether shipping data plus code alongside weights becomes the default for Chinese open-source teams.