Attempted AWQ quantization using calibration dataset. Added placeholder 1 token images to the text calibration dataset. Took some messing around and might look weird; I needed to patch vLLM for support even though it is trying to support Qwen3.5 out of the gate.

It's ... weird-shaped currently due to unpacking and might not be useful to anyone, lol. Seems slower than GGUF. Unsure how well it works. However, I spent a good few hours of my computer's time on making it happen, so releasing regardless.

Downloads last month
24
Safetensors
Model size
36B params
Tensor type
I64
·
I32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Lambent/Qwen3.5-35B-A3B-Base-AWQ-W4A16

Quantized
(9)
this model

Datasets used to train Lambent/Qwen3.5-35B-A3B-Base-AWQ-W4A16