ZGCM-1: fully open 7B claims to rival 235B on math, search

ZGCM-1, a 7-billion-parameter dense model trained from scratch, holds its own against systems its authors describe as "orders of magnitude larger" on math and agentic-search benchmarks. In the abstract, the team writes that 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." That is a claim of competitiveness, not a benchmark table. The abstract publishes no per-suite scores.

The design bet is stated up front. The authors argue that "compact models cannot passively memorize the open web, but can overcome parametric capacity limits by coupling deliberate internal thinking with active external tool use." To get there they combine interleaved gated sliding-window and full attention, an FP8 Muon optimizer, and a progressive curriculum that scales context from 16K to 64K to 256K tokens. They report roughly a 4.2x improvement in pre-training time-to-loss at the 16K stage, and reformulate interaction traces as Markov Decision Processes for the agentic post-training.

What separates this from most efficiency claims is the release plan. The team says it will publish weights from pre-training, mid-training and post-training, intermediate checkpoints, the training code and implementation details, per-stage data recipes, and Weights & Biases logs documenting the training process. That is enough surface area for outsiders to rerun the pipeline rather than trust the abstract.