---
title: Genoria · Qwen3.6-27B Fable-Fusion-711
emoji: 🔮
colorFrom: purple
colorTo: gray
sdk: docker
app_port: 7860
license: apache-2.0
short_description: Chat + vision demo · DavidAU Qwen3.6-27B Fable-Fusion-711
suggested_hardware: l40sx1
startup_duration_timeout: 1h
thumbnail: https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF/resolve/main/valhalla.webp
models:
  - DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF
tags:
  - llama.cpp
  - gguf
  - qwen3.6
  - vision
  - thinking
---

# Genoria · Qwen3.6-27B Fable-Fusion-711 demo

Live chat demo of [DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF](https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF) —
a multi-stage Qwen3.6-27B fine-tune (thinking/reasoning model, 256k context,
vision-capable) served by [llama.cpp](https://github.com/ggml-org/llama.cpp)'s
`llama-server`, using its built-in web UI. Image upload works out of the box
(the Space loads the repo's `mmproj-F16.gguf`).

The same endpoint speaks the **OpenAI API**:

```bash
curl https://<this-space-host>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Hello!"}]}'
```

## What's running

| Setting | Value |
|---|---|
| Quant | `…NEO-IQ2_M.gguf` (11.7 GB, regular / non-MTP, NEO-MAX imatrix) |
| Context | 8,192 tokens (model card minimum; native is 262,144) |
| KV cache | q8_0 |
| Samplers | temp 1.0 · top_p 0.95 · top_k 20 · min_p 0 · rep_pen 1.0 (model-card "thinking mode" defaults) |

**Speed expectations:** on the free `cpu-basic` tier (2 vCPU / 16 GB) a dense
27B runs at well under 1 token/s — enough to verify the model responds, not to
converse. Duplicate the Space onto GPU hardware (L40S or better) and it becomes
a real-time demo.

## Running it on GPU

1. Settings → switch hardware to `l40s` (48 GB) or bigger.
2. Add Space **variables** (they're passed through as Docker build args and
   runtime env — a rebuild picks them up):
   - `LLAMACPP_IMAGE` = `ghcr.io/ggml-org/llama.cpp:full-cuda`
   - `MODEL_FILE` = `Qwen3.6-27B-Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-Q6_K.gguf` (23.6 GB) — or any other quant from the repo
   - `CTX_SIZE` = `32768` (or higher, VRAM permitting)

All entrypoint knobs (`TEMP`, `TOP_P`, `TOP_K`, `MIN_P`, `REPEAT_PENALTY`,
`MAX_TOKENS`, `THREADS`, `NGL`, `KV_QUANT`, `ENABLE_VISION`, `EXTRA_ARGS`) are
plain runtime env vars — settable as Space variables without editing code.

## About the MTP quants

The source repo ships both **regular** and **MTP** (multi-token-prediction)
quants. Per DavidAU's card: MTP quants can exceed regular quants' speed when
token acceptance ≥ 50%, but want `temp ≤ 1` and `rep_pen = 1`, and regular
quants win for creative/high-temp use. This Space defaults to a regular quant
for maximum compatibility; to try an MTP file set `MODEL_FILE` to one of the
`…MTP-*.gguf` names.

## Credits

- Model: [DavidAU](https://huggingface.co/DavidAU) and collaborators (Apache-2.0)
- Runtime: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) server + web UI
- Space by Genoria