ZGCM-1 is a fully open 7B dense model whose training cluster was operated by agent swarms

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2609.13356, submitted 2026-09-11 with 227 HuggingFace upvotes, releases a 7B dense model trained from scratch on the premise that small models cannot memorize the web but can trade parametric capacity for deliberate thinking plus external tool use. The recipe is fully open: interleaved gated sliding-window and full attention, an FP8 Muon optimizer, and a progressive curriculum scaling context through 16K, 64K and 256K while reformulating interaction traces as MDPs. On math reasoning and agentic search it stays competitive with Qwen3-235B-A22B and GLM-5.1, and the authors report agent swarms autonomously handling cluster operations, data curation and diagnostic evaluation during the run.