🍷 Winery Qwen3.8 27B · Grand Cru
The biggest dense bottle in the cellar: a weighted soup of Qwen3.8-27B and six of its best fine-tunes: Swift 1.5 and Swift 1.0, Qwopus Flash, two Fable-5 distills and Huihui's abliterated build. Every tensor was streamed straight from the donors' bf16 weights and averaged in fp32, then quantised once to Q8_0. 28.6 GB.
Tasting notes
Winery quick eval (0-shot, thinking off; same harness as every Winery model):
| MMLU | ARC-Challenge | GSM8K | Avg | |
|---|---|---|---|---|
| 27B Grand Cru | 76.5 | 98.0 | 81.7 | 85.4 |
| 9B Assemblage | 73.0 | 97.3 | 88.3 | 86.2 |
| 9B Grand Cru | 70.0 | 96.7 | 90.0 | 85.6 |
Grapes (linear soup, fp32 from bf16)
All 850 language-model tensors line up across the seven checkpoints. The MTP head is dropped. Config in recipe.json.
Use
Turn thinking off for straight answers (chat_template_kwargs: {"enable_thinking": false}).
llama-server -hf WineryLabs/Winery-Qwen3.8-27B-GrandCru-GGUF -c 32768
ollama run hf.co/WineryLabs/Winery-Qwen3.8-27B-GrandCru-GGUF
Needs about 30 GB of VRAM + RAM combined; llama.cpp's --fit spreads it across a 16 GB GPU and system RAM. Works in llama.cpp (recent builds with Qwen3.8 support), Jan, LM Studio and the Winery app.
Licence
The Swift weights carry the Swift Open License v1.0 (LICENSE); Qwen3.8-27B and the other donors are Apache 2.0 (LICENSE-APACHE-2.0). Attribution in NOTICE.
Copyright 2026 UkisAI. Swift Contribution licensed under the Swift Open License v1.0 (https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b/blob/main/LICENSE). Derivative of Qwen3.8-27B, Copyright 2026 Alibaba Cloud, Apache License 2.0.