Researchers release ZGCM-1, a 7B open model for math and agentic search with 256K context and competitive performance

Researchers introduced ZGCM-1, a fully open 7B dense foundation model trained from scratch with an emphasis on extreme data, system, and algorithmic efficiency. The core premise is that compact models cannot rely on passive memorization of the open web but can overcome parametric capacity limits by pairing deliberate internal reasoning with active external tool use.

To operationalize this paradigm across a 256K context window, the team developed an end-to-end, high-efficiency training recipe that includes architecture and system co-design—interleaved gated sliding-window and full attention, and a stable FP8 Muon optimizer—and a progressive curriculum with Markov Decision Process mid-training. Context scaling was staged from 16K to 64K and then to 256K.

The authors also established an AI-native R&D workflow in which agent swarms autonomously manage cluster operations, data curation, and rapid diagnostic evaluation, aiming to accelerate iteration and reproducibility.

Extensive evaluations indicate that ZGCM-1-7B is competitive within the 7B model family on general benchmarks. On several challenging mathematical reasoning and agentic search suites, it remains competitive with frontier models orders of magnitude larger, such as Qwen3-235B-A22B and GLM-5.1.

The team reports that their pre-training design delivers approximately a 4.2x efficiency improvement in 16K pre-training time-to-loss. Across the full development lifecycle, they distill eight actionable empirical findings spanning architectural scaling, supervised fine-tuning quality pruning, long-context generalization, and agentic co-training dynamics.

To support community research, the authors open-source model weights from pre-training, mid-training, and post-training stages, along with intermediate checkpoints, training code, per-stage data and data recipes, and Weights & Biases logs.