user3261928958/ZGCM-1-7B-Demo

ZGCM-1-7B Demo

A demo page for zgcagi/ZGCM-1-7B, a fully open 7.39B dense foundation model from Zhongguancun Academy, trained from scratch for mathematical reasoning and agentic search (256K context, thinking + direct-answer modes in one checkpoint).

• —MATH-500 97.13 · AIME 2026 75.00 · HMMT 2025 70.42
• —Tech report: https://arxiv.org/abs/2609.13356
• —Training code: https://github.com/zgcagi/ZGCM-1

Current state: Static (no GPU yet)

This Space currently ships as a Static Space, because ZeroGPU requires either a PRO subscription or an account older than 30 days.

app.py + requirements.txt in this repo are the ready-to-run Gradio app targeting ZeroGPU (zero-a10g, 48 GB slice). Once hardware is available, flip the frontmatter:

sdk: gradio           # was: static
app_file: app.py      # was: index.html

No other change is needed — the app already follows the ZeroGPU contract:

• —import spaces is the first import (ZeroGPU monkey-patches torch.cuda.*).
• —The model is loaded at module scope with .to("cuda"), so ZeroGPU streams the weights into VRAM on the first @spaces.GPU call.
• —trust_remote_code=True — ZGCM-1 ships its own modeling_zgcm.py / configuration_zgcm.py.
• —The checkpoint is ~15 GB in bfloat16 and only depends on torch + transformers (no flash-attn), so it fits the 48 GB ZeroGPU slice.