🍷 Winery Qwen3.8-Flash-Next · Cuvée
One blend of Qwen3.8-Flash-Next (125B MoE, 512 experts, 10 active) in three bottles, sized so it runs on a gaming PC with Winery Strata. That's a fork of Strata with a wine-red glass UI, where Cuvée is the first model in the installer.
- Short thinking from Swift 1.5 (UkisAI)
- Refusals poured out by Huihui's abliteration
- Base everywhere else: routed experts' gate/up, routers, hyper-connections, PLE n-gram table
🍷 Visit the Cuvée Space for the expert-cellar map and a bottle picker for your RAM.
The bottles
| Folder | For | Experts | Expert types | Size |
|---|---|---|---|---|
POCKET/ | 8 GB GPU · 24-32 GB RAM | 256 per layer | IQ2_XS gate/up · Q2_0 down, lean dense (Q4_K attn) | ~48 GB |
SLIM/ | 32 GB RAM | 256 per layer | IQ3_XXS gate/up · IQ4_NL down | ~60 GB |
RESERVE/ | 64 GB RAM | all 512 | IQ3_XXS gate/up · IQ4_NL down | ~86 GB |
GRAND/ | 96-128 GB+ RAM | all 512 | Q4_K gate/up · Q5_1 down | ~113 GB |
All three keep the dense weights in Q6_K/Q8_0 and the 28.8 GB PLE table in IQ4_NL (Strata reads it from the SSD). They were quantised with Unsloth's imatrix for Flash-Next.
Slim drops half the experts, the same move as ISTA-DASLab's Coder. The difference is that it keeps the 256 experts routed most often on general text (Unsloth's calibration routing counts, 72-91% of routed tokens per layer, 77% on average), not the ones code uses. Expect it to be a little weaker than Reserve, mostly on rare topics.
How it was made
Comparing the fine-tunes with the base tensor by tensor showed they touch different parts of the model:
| Tensors | Swift 1.5 | Huihui |
|---|---|---|
| attention q/k/v · DeltaNet qkv/z · shared-expert gate/up | ✓ | |
| attention/DeltaNet output · shared-expert down | ✓ | ✓ |
| routed experts' down | ✓ | |
| everything else |
So Cuvée = base + (Swift − base) + (Huihui − base) (task arithmetic, FP32 per tensor). Every tensor was streamed straight from the three BF16 checkpoints with HTTP range requests, merged and written to a Q8_0 GGUF. The Q8_0 was then quantised once to each bottle. The vision tower and MTP head are the original model's (Strata fetches them itself).
Run it
With Winery Strata (NVIDIA RTX 20-50 or a recent AMD card; 8 GB cards such as the RTX 4060 laptop GPU use POCKET, which setup picks for them by itself):
git clone https://huggingface.co/WineryLabs/Winery-Strata && cd Winery-Strata
START-HERE.bat --setup --family winery --model POCKET # Windows, 8 GB GPU; SLIM / RESERVE / GRAND for more
START-HERE.bat --setup --family winery --model SLIM # Windows; RESERVE / GRAND for more RAM
./setup.sh --setup --family winery --model RESERVE # Linux
With llama.cpp (a build with qwen4exp support), e.g. llama-server -m RESERVE/Winery-Cuvee-RESERVE-00001-of-00002.gguf -c 32768 -ngl 99 --n-cpu-moe 48.
Not measured yet: speed on real hardware, and benchmark scores against the parent models. Sanity chats on CPU are in
TASTING.md.
Licence
A derivative of Qwen3.8-Flash-Next (Qwen Community License 1.0, LICENSE-QWEN) that contains UkisAI's
Swift Contribution (Swift Open License v1.0, LICENSE: free, including commercial use below US$1M yearly
revenue). Huihui's weights are under the Qwen licence. Attribution and our changes are in NOTICE.