CoolFace
Apppublic

kf3paaw/qwen3-8-27b-obliterated-demo

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
0likes
App README

🔓 Qwen3.8-27B — OBLITERATED (Demo)

Interactive chat UI for **OBLITERATUS/Qwen3.8-27B-OBLITERATED**, an "abliterated" (refusal-removed) chat model built on Qwen3.8-27B. Per the model card it keeps near-stock capability (MMLU ~83.7%, 7/8 advanced real-world tasks) with zero refusals on both thinking modes.

⚠️ Important: this Space does NOT call Hugging Face Serverless

Verified limitation. Hugging Face's hosted Serverless Inference does not serve this model. A real, authenticated call returns:

BadRequestError: The requested model 'OBLITERATUS/Qwen3.8-27B-OBLITERATED'
is not supported by any provider you have enabled.  (code: model_not_supported)

So this Static Space cannot run the 27B weights by itself. Instead, the browser-only UI below talks to an OpenAI-compatible `/v1/chat/completions` endpoint that YOU provide — i.e. a server that actually hosts the weights. No token is ever embedded in the page; your key (if any) lives only on that endpoint.

How to use it

  1. 1.Deploy/host the model somewhere that exposes OpenAI-compatible /v1/chat/completions. Easiest options:
  2. 2.Ollama (local or remote): ollama run OBLITERATUS/Qwen3.8-27B-OBLITERATED then use http://<host>:11434/v1/chat/completions.
  3. 3.llama.cpp / llama-cpp-python server: --api-key + /v1/chat/completions.
  4. 4.vLLM: vllm serve OBLITERATUS/Qwen3.8-27B-OBLITERATED --served-model-name OBLITERATUS/Qwen3.8-27B-OBLITERATED.
  5. 5.A Cloudflare Worker / reverse proxy that holds your HF (or other) token server-side and forwards to your own hosted instance.
  6. 6.Open the Space, paste that endpoint URL into the "Model endpoint" box.
  7. 7.Chat. The UI sends the model card's optimal settings: temperature=0, repetition_penalty=1.15, thinking OFF, no system prompt, max_tokens (slider, default 2048).

Notes

  • —Keep thinking disabled when self-hosting — the card warns thinking can re-derive refusals.
  • —The UI accepts bare origins, /v1, /v1/chat, or full /v1/chat/completions URLs and normalizes them.
  • —This is a Static Space (no server, no PRO needed): just index.html.