huguforever/zerogpu-fast

🚀 Fleet · fast

ZeroGPU worker serving [huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2](https://huggingface.co/huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2) (Apache-2.0) — OpenAI gpt-oss-20b MoE (21B/3.6B active) in its native MXFP4 checkpoint (~14GB), SFT'd by huihui-ai on abliterated samples.

Role in the fleet

Fast general worker: cheapest/quickset turn for summarising, drafting, routing, classification. Text-only — no tool calling on this endpoint (gpt-oss speaks the harmony format; give tool-needing tasks to reason/code workers or the CEO).

Call it

from gradio_client import Client

c = Client("huguforever/zerogpu-fast", hf_token="hf_...")  # token bills ZeroGPU quota to you
res = c.predict(
    messages='[{"role":"user","content":"Summarize this diff in 3 bullets"}]',
    tools='[]', max_tokens=2048, temperature=0.6,
    api_name="/chat",
)
assistant = res["message"]

Activation checklist

1. 1.Space Settings → Hardware → ZeroGPU (size large, 48GB — MXFP4 runs natively on the Blackwell RTX Pro 6000).
2. 2.Add HF_TOKEN secret so gradio_client calls bill your account.
3. 3.Lightest build of the fleet (~14GB download) — fast first boot.

⚠️

Abliterated model — refusal behaviour removed; not aligned. For research use.