Qwen3.8-27B-GSQ-INT2-10e

These are research artifacts, not intended for practical use. I made these models to test GSQ with different hyperparameters. I trained them with limited resources, and they significantly underperform the original model, but they are not broken.

Hyperparameter	Value
Base model	Qwen/Qwen3.8-27B
Method	GSQ
Weight dtype / groups	INT2 (group 32)
Epochs per layer	10
LM head / embeddings	INT8 / INT8 (group 64)
Retained floating dtype	bfloat16
Training logits dtype	float32
Logits / scale learning rate	0.0001 / 5e-05
Calibration mixture	reasoning-balanced
Calibration / GPTQ / reconstruction samples	1024 / 512 / 64
Sequence length	4096
Batch / microbatch	16 / 16
Seed	42
Format	Safetensors; compressed-tensors (pack-quantized)
MTP / vision	Original BF16 MTP preserved; vision excluded