Downloads · 30 days
431
49% of all-time downloads
r0b0tlab/Qwen3.8-27B-EXL3-4.00bpw
Qwen3.8-27B-EXL3-4.00bpw is a image-text-to-text model from r0b0tlab. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for exllamav3. The card lists the license as apache-2.0.
EXL3 (ExLlamaV3) quantization of Qwen/Qwen3.8-27B, tuned for a 24 GB RTX 3090 at the model's full 262,144-token context.
Downloads · 30 days
431
49% of all-time downloads
All-time downloads
872
Public
Parameters
8.3B
16.5 GB on disk
Likes
3
Public
Click a slice to open those files.
.safetensors16.5 GB · 100%
How the weights are stored.
I166.8B · 83%
From the Hugging Face model README
EXL3 (ExLlamaV3) quantization of Qwen/Qwen3.8-27B, tuned for a 24 GB RTX 3090 at the model's full 262,144-token context.
Conversion: ExLlamaV3 v1.5.0, single pass with
convert.py -b 4.00 -vb 6 -mb 4 -hb 6 — 4.00 bpw decoder, 6 bpw lm_head, vision tower
6 bpw, native MTP head 4 bpw. Serve with r0b0tlab/exllamav3
branch community @ 355c6ee (native DFlash2 CUDA; no dflash2-pathway overlay).
Pair with the DFlash2 draft: r0b0tlab/Qwen3.8-27B-DFlash2-EXL3-4.00bpw.
ExLlamaV3 (fork clone, branch community):
python examples/chat.py -m <this-dir> -mode chatml \
-dm <draft-dir> -cs 262144 -cq 3
Container (click-run) — auto-downloads this repo on first run:
docker run --gpus all -v qwen38-models:/models \
ghcr.io/r0b0tlab/qwen38-exl3-dflash2:1.5.0-native
Full metrics, harnesses and the served model: https://github.com/r0b0tlab/qwen38-exl3-dflash2