Evaluation
Thinking mode, pass@1 averaged over 1–32 seeds, same settings for all models.
Usage
Runs on vLLM only (SGLang does not support this 3-bit format yet), on NVIDIA H100, A100, L40S, RTX Pro 6000 and B200 GPUs.
vLLM:
vllm serve qda-lab/QDA-Qwen3.8-27B-W3A16
How it was made
Built on ISTA-DASLab's public 3-bit GSQ checkpoint of Qwen3.8-27B: its 3-bit integer weights (groups of 128) are kept unchanged,
and QDA re-tuned the group scales and some BF16 tensors; the output head and token embeddings are 4-bit. The recipe was chosen by QDA's automated search.
Reproduce
hf download qda-lab/qda-recipe-inputs --repo-type dataset --include "QDA-Qwen3.8-27B-W3A16/*" --local-dir inputsqda build inputs/QDA-Qwen3.8-27B-W3A16/qwen3.8-27b-w3a16.yaml --set "devices=[0,1]"
Needs the QDA library (the qda command). Recipe: qwen3.8-27b-w3a16.yaml; inputs
(the 3-bit start checkpoint in start/, 14.6 GB, and the distillation data): QDA-Qwen3.8-27B-W3A16/ in qda-lab/qda-recipe-inputs.
Takes about 4.5 h on 2× H100 (measured). The rebuild matches this checkpoint's KL to BF16 within 5 %; weights are not bit-identical.
The recipe starts from the 3-bit GSQ checkpoint ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ by IST Austria's DAS Lab (Apache-2.0), with its embedding dequantized to BF16 and Qwen3.8-27B's BF16 MTP block added; the dataset folder holds that start (start/, with ISTA's licence and our changes in its README) and derive_start.py, which rebuilds it byte for byte from ISTA's release:
hf download ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ --revision b5ce0b76f60020a875dee4f6ec9d934cca4121e4 --local-dir istahf download Qwen/Qwen3.8-27B --revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 --include "model.safetensors.index.json" "model-00018-of-00018.safetensors" --local-dir qwenpython inputs/QDA-Qwen3.8-27B-W3A16/derive_start.py ista qwen start
License
Apache-2.0, the same as the base model.
Citation
TBD.