huguforever/zerogpu-code

💻 Fleet · code

ZeroGPU worker serving [huihui-ai/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated](https://huggingface.co/huihui-ai/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated) (Apache-2.0) — the agentic coding MoE (30.5B/3.3B active, 256K ctx), loaded 4-bit NF4. Non-thinking by design: built for tool-calling agent loops.

Role in the fleet

Coding worker for OpenClaw agents/subagents: file edits, codegen, repo tasks. Fastest per token of the Qwen workers; tool-calling enabled via <tool_call> markers.

Call it

from gradio_client import Client

c = Client("huguforever/zerogpu-code", hf_token="hf_...")  # token bills ZeroGPU quota to you
res = c.predict(
    messages='[{"role":"user","content":"Write pytest tests for utils.py"}]',
    tools='[]', max_tokens=2048, temperature=0.6,
    api_name="/chat",
)
assistant = res["message"]

messages and tools are JSON strings (OpenAI format). Qwen-family <tool_call> markers are parsed into OpenAI tool_calls. Streaming is not exposed on the Space — use the local fleet translator for an OpenAI /v1 view with SSE.

Activation checklist

1. 1.Space Settings → Hardware → ZeroGPU (size large, 48GB).
2. 2.Add HF_TOKEN secret so gradio_client calls bill your account.
3. 3.First build downloads ~61GB of bf16 weights and quantizes to NF4 — allow the 1h startup timeout.
4. 4.Test one call before enabling other fleet Spaces.

⚠️

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