🍷 Winery Qwen3.5 9B · Atelier
A web design + 3D specialist: a region-weighted merge of six Qwen3.5-9B web and coding fine-tunes, built with the Winery fusion compiler. Give it a brief and it writes a complete single-file site: bold landing pages, glassmorphism, scroll animations and live three.js scenes. 9.5 GB at Q8_0.
👉 Atelier's own Space was designed and coded by this model.
Grapes
| Donor | Why it's in |
|---|---|
| Tesslate/OmniCoder-9B | front-end / UI generation |
| smirki/tesslate-sft-9b-v3 | Tesslate UI SFT checkpoint |
| OrionLLM/OxCoder-9B | general coding |
| Jackrong/Qwopus3.5-9B-Coder | Opus-style coding |
| YangLiu258/Qwen3.5-9B-Exp09-DPO-Html0708 | HTML preference-tuned |
| armand0e/Qwen3.5-9B-Coder | coding |
Embeddings and early layers lean on the UI tunes, later layers get more of the coders. All 427 tensors were mixed with zero fallbacks; full recipe in recipe.txt.
Showcase
Unedited one-shot outputs (thinking off, temp 0.6), in showcase/:
| 3D hero (three.js) | Landing page |
|---|---|
 |  |
Picked over two sibling blends because it was the only one whose three.js scene rendered without errors. It hasn't been benchmarked on MMLU-style tests; it's tuned for front-end work, not trivia.
Use
Turn thinking off for straight code output (chat_template_kwargs: {"enable_thinking": false}) and give it room: full pages run 2-8k tokens.
llama-server -hf WineryLabs/Winery-Qwen3.5-9B-Atelier-GGUF -c 32768
ollama run hf.co/WineryLabs/Winery-Qwen3.5-9B-Atelier-GGUF
Works in llama.cpp (recent builds with Qwen3.5 support), Jan, LM Studio and the Winery app.