Beijing Zhongguancun Academy

Seven PhD students at Beijing Zhongguancun Academy have open-sourced ZGCM-1, a 7B model trained from scratch over a summer. The release includes stage-by-stage training data and recipes, weights, training code, intermediate checkpoints and logs; no licence was specified.

ZGCM-1 uses local plus global attention, with context extended from 16K to 256K. The team reports roughly 3.94× throughput at 256K, KV cache about one-sixth the size, and about 4.2× pretraining efficiency at 16K versus the BF16/AdamW baseline. They also used hundreds of agents for data processing, experiments, log review and evaluation, rating agent autonomy on an L1–L5 scale.