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-Qwen3.6-35B-A3B-W3A16
How it was made
Weights are 3-bit integers in groups of 128, with the output head and token embeddings at 4 bits; the group scales
were then fine-tuned by distillation. MoE routers and shared experts stay in BF16. The recipe was chosen by QDA's automated search.
Reproduce
hf download qda-lab/qda-recipe-inputs --repo-type dataset --include "QDA-Qwen3.6-35B-A3B-W3A16/*" --local-dir inputscd inputs/QDA-Qwen3.6-35B-A3B-W3A16qda build qwen3.6-35b-a3b-w3a16-ace.yaml --set "devices=[0,1]" --set residency=streamqda build qwen3.6-35b-a3b-w3a16-h.yaml --set "devices=[0,1]" --set residency=streamqda build qwen3.6-35b-a3b-w3a16-i.yaml --set "devices=[0,1]" --set residency=streamqda build qwen3.6-35b-a3b-w3a16-soup.yaml
Needs the QDA library (the qda command). Run the four recipes from inside the folder, in this order: each reads the
previous outputs (./start, ./w3-E, ./w3-H, ./w3-I). Recipes and inputs (the 3-bit start checkpoint in start/,
17.3 GB, four distillation trace sets and the default calibration text):
QDA-Qwen3.6-35B-A3B-W3A16/ in qda-lab/qda-recipe-inputs.
Takes about 5.5 h on 2× H100 (measured: -ace 3.4 h, -h 1.05 h, -i 0.9 h; -soup runs on CPU). The rebuild matches
this checkpoint's KL to BF16 within 5 %; weights are not bit-identical. -i uses sdpa attention where the original
build used eager, so that it fits 80 GB cards; the maths is the same, the bytes are not.
The recipe starts from a 3-bit checkpoint we quantized from Qwen3.6-35B-A3B with the official GSQ code
(IST-DASLab/GSQ, paper); the dataset folder holds it (start/).
License
Apache-2.0, the same as the base model.
Citation
TBD.