EXL3 quantization of ukisai/Swift-Qwen3.8-27b, UkisAI's reasoning-efficient derivative of Qwen3.8-27B, using 58.3% fewer thinking tokens while maintaining near-identical performance (<1% loss) and as a result getting a x1.95 speed-up on several tasks. 13.9 GB of weights — sized to leave room for large KV caches on 24 GB GPUs: with the built-in MTP draft head the full native 262k context fits fully resident.
Quantization details
| setting | value |
|---|---|
| format | EXL3 v1.5.0 (quantization_config.json) |
| average bits | 3.5 bpw (module-adaptive) |
| head bits | 6 (default) |
high-quality mode (-hq) |
no |
The conversion was done with this simple command:
python3 convert.py \
-i /mnt/d/models/ukisai/Swift-Qwen3.8-27B \
-o /mnt/d/models/charnet3d/Swift-Qwen3.8-27B-Exl3-3.5bpw \
-w /mnt/d/temp_scratchpad \
-b 3.5
This is shared as-is to save time for others who need it.
No accuracy testing or quality benchmarks were done, except normal usage which seemed fine.
The creator is not liable if this model nuked your house or deleted your prod db.
Requirements
- exllamav3 — the
qwen3_5architecture used here is supported upstream, including MTP speculative decoding; no fork required for basic inference on x86 CUDA GPUs.
License
swift-open-license-1.0, inherited from the base model.
Credits
- Base model: ukisai/Swift-Qwen3.8-27b
- Quantization format & tooling: exllamav3 (turboderp)
- Readme formatting inspired from Mia-AiLab/Qwen3.8-27B-EXL3-3.5bpw
- Downloads last month
- 157