ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search
ZGCM-1 is a 7B open model built around long-context reasoning and external tool use, with the authors releasing the full training stack.
The paper says the model was trained from scratch with a 256K context recipe, using sliding-window and full attention plus an FP8 Muon optimizer. Its training includes progressive context scaling and mid-training that reformulates interaction traces as Markov Decision Processes. The authors report competitive results among 7B models, and say it holds up on some math and agentic search benchmarks against much larger frontier models. They also claim a roughly 4.2x improvement in 16K pre-training time-to-loss and release weights, checkpoints, code, data recipes, and logs. HF Daily Papers' note
The paper says the model was trained from scratch with a 256K context recipe, using sliding-window and full attention plus an FP8 Muon optimizer. Its training includes progressive context scaling and mid-training that reformulates interaction traces as Markov Decision Processes. The authors report competitive results among 7B models, and say it holds up on some math and agentic search benchmarks against much larger frontier models. They also claim a roughly 4.2x improvement in 16K pre-training time-to-loss and release weights, checkpoints, code, data recipes, and logs. HF Daily Papers' note
score 7