huguforever/zg-reasoner-qwen38-27b

🧠 Fleet reasoning worker — Qwen3.8-27B (abliterated)

Reasoning worker for agents/subagents: dense 27B with thinking mode (`reasoning_effort` low/medium/xhigh) — the model to call when a spawned agent must reason hard, plan, or analyze before answering.

• —Model: huihui-ai/Huihui-Qwen3.8-27B-abliterated — Apache-2.0, 27.8B dense, 262K ctx, hybrid attention, vision tower (text-only path exposed)
• —Runtime: 4-bit NF4 on ZeroGPU large (48 GB) — 1× quota per call
• —Endpoint: Gradio API /chat (streaming, multi-turn)

Activation (one-time, in this Space's Settings tab)

1. 1.Hardware → ZeroGPU (large = 48 GB, default).
2. 2.Model weights preload from the Hub during the build (preload_from_hub), so the first call is fast.

ZeroGPU quota is account-wide and shared with your other ZeroGPU Spaces (PRO: 40 min/day included, then $1/10 min of credits).

Calling from an agent (gradio_client)

from gradio_client import Client

client = Client("huguforever/zg-reasoner-qwen38-27b")  # add hf_token=... for private
reply = client.predict(
    "Plan the refactor of module A into three steps.",  # message
    1024,      # max_new_tokens
    0.7,       # temperature
    0.9,       # top_p
    True,      # enable_thinking
    "high",    # reasoning_effort
    api_name="/chat",
)
print(reply)

Reuse the same Client instance to keep conversation history.

⚠️ Abliterated model: refusal/safety behaviour is substantially reduced. Outputs may be inaccurate, sensitive or inappropriate.