The researchers behind ZGCM-1 have released a 7-billion-parameter foundation model alongside unusually broad training artifacts. The package includes weights from pre-training, mid-training and post-training, intermediate checkpoints, code, stage-by-stage data and recipes, and experiment logs.
ZGCM-1 is designed around the idea that a compact model can offset limited memorized knowledge through deliberate reasoning and external tool use. Its architecture mixes sliding-window and full attention, supports a context window up to 256,000 tokens, and was trained with a curriculum that expanded context from 16,000 to 64,000 and then 256,000 tokens.
The authors report that the model is competitive with other 7B systems on general benchmarks and with much larger models on some mathematical-reasoning and agentic-search tests. They also claim roughly a 4.2-fold improvement in time-to-loss during 16K pre-training from their system and training design.
Those performance figures come from the project’s own evaluations and the paper is a new arXiv preprint, not a peer-reviewed result. The more immediately verifiable contribution is the release itself, which gives researchers access to intermediate training stages and recipes that are often omitted from open-weight model launches.