ZGCM-1

Zhongguancun Academy's first language model and its most complete release: a 7B dense model pretrained from scratch on 4.19T tokens under MIT, published with the pre-training checkpoint, the staged 16K, 64K, and 256K mid-training checkpoints, and a technical report from the ZGCM Team (Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence). The premise is that a compact model cannot memorise the open web but can compensate by pairing deliberate internal thinking with external tool use, so the recipe targets math reasoning and agentic search across a 256K context: interleaved gated sliding-window and full attention, a stable FP8 Muon optimizer, a progressive context curriculum, and a mid-training stage that reformulates interaction traces as Markov decision processes. The report also describes an "AI-native" R&D workflow in which agent swarms run cluster operations, data curation, and diagnostic evaluation, and claims a roughly 4.2x pre-training efficiency gain from the architecture and system co-design.

Self-reported results at 7B: MATH-500 97.13, AIME 2026 75.00, HMMT 2025 70.42, with the best average rank across the report's 14 reasoning benchmarks among seven compared 7B models; with tools, WebWalkerQA 63.09 and BrowseComp 19.43. The report positions it as competitive on those suites with far larger open models such as Qwen3-235B-A22B and GLM-5.1. Not yet scored by Artificial Analysis.