ZGCM-1: fully open 7B model matches much larger frontier models on math and agentic search
_reachsumit · x · 2026-09-15
ZGCM-1 is a fully open, from-scratch 7B dense foundation model built for math reasoning and agentic search with a 256K context, premised on the idea that compact models beat capacity limits via deliberate internal thinking plus active tool use.
Key techniques:
- Interleaved gated sliding-window and full attention, with a stable FP8 Muon optimizer
- Progressive curriculum: context scaling across 16K/64K/256K, and reformulating interaction traces as MDPs for mid-training
- AI-native R&D workflow where agent swarms autonomously manage cluster ops, data curation, and rapid diagnostic evals
On several challenging math and agentic search suites, ZGCM-1-7B is competitive with models orders of magnitude larger, such as Qwen3-235B.
Related event: 7 PhD Students Train Open-Source 7B Model ZGCM-1 in One Summer(3 posts)→
More from Models
- AWS's open-source Strands Harness claims 45% cheaper than Claude Code, reigniting harness debate — blaizedsouza · 2026-11-02
- Yacine calls Terra High 'kind of an insane model' — yacineMTB · 2026-11-02
- Running DeepSeek V4.1 on a single H200 at $5.30/hr — daluoseo · 2026-11-02
- Leaked Fable 5.5 checkpoint video called 'pure code' — 'will change the game' — kimmonismus · 2026-11-02
- Claude Opus 5.5 quality complaints spark debate on why continual learning is so hard — JeremyNguyenPhD · 2026-11-02
- Why is AI sexist? Reddit debate on models overlooking Iceland's women's strike — Mathemodel · 2026-11-02