ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
The ZGCM-1 7B model matches the math and agentic search performance of 235B-parameter frontier models by substituting raw parametric memorization with internal reasoning and tool-use over a 256K context, built on a training recipe that is 4.2 times more efficient. This enables you to migrate complex, long-context agentic search pipelines from costly proprietary APIs to local, cheap 7B model deployments without a drop in reasoning capabilities. Since the entire stack—including intermediate checkpoints and data recipes—is fully open-source, you can easily customize and fine-tune the model's interaction traces for your specific tool suite.
ZGCM-1 is a fully open 7B dense model with 256K context, released with weights across training stages, code, data recipes, checkpoints, and logs. The practical shift is that small-model agentic search and math systems can be reproduced and adapted end-to-end, with a reported ~4.2x improvement in 16K pretraining time-to-loss and competitive results against much larger frontier models on targeted reasoning/search tasks.