🍷📖 Winery Qwen3.5 9B · Fable Reserve
A soup of five Qwen3.5-9B fine-tunes distilled from Claude Fable 5, merged in full bf16 precision (not on quantised weights) and then quantised once to Q8_0. It reasons cleanly, writes well and builds good-looking front-ends: landing pages, three.js scenes and single-file HTML. 9.5 GB.
👉 Fable Reserve's own Space was designed and coded by this model.
Tasting notes
Winery quick eval (0-shot, thinking off; same harness as every Winery model):
| MMLU | ARC-Challenge | GSM8K | Avg | |
|---|---|---|---|---|
| Fable Reserve | 70.0 | 95.3 | 75.0 | 80.1 |
| Qwable-9B (anchor donor, solo) | 71.0 | 94.7 | 63.3 | 76.3 |
| 9B Assemblage | 73.0 | 97.3 | 88.3 | 86.2 |
| 9B Grand Cru | 70.0 | 96.7 | 90.0 | 85.6 |
| Qwen3.5-9B (stock) | 77.0 |
Grapes (linear soup, bf16)
Only the language model and LM head are merged; the vision tower comes from stock Qwen3.5-9B. Config in recipe.json.
Showcase
Unedited one-shot outputs (thinking off, temp 0.6), in showcase/:
| Landing page | three.js scene |
|---|---|
Use
Turn thinking off for straight answers or code (chat_template_kwargs: {"enable_thinking": false}).
llama-server -hf WineryLabs/Winery-Qwen3.5-9B-Fable-Reserve-GGUF -c 32768
ollama run hf.co/WineryLabs/Winery-Qwen3.5-9B-Fable-Reserve-GGUF
Works in llama.cpp (recent builds with Qwen3.5 support), Jan, LM Studio and the Winery app.