int4 quantization of Qwen/Qwen3.8-27B, produced with compressed-tensors by streaming the checkpoint tensor-by-tensor (the model is never fully instantiated).

4-bit integer weights in groups of 128 with a BF16 scale. Smallest of the set and the highest error.

All quantizations of this model

Variant	Format	Size	vs BF16	Mean rel. error	Linears quantized	Left BF16
Qwen3.8-27B-FP8	`float-quantized`	30.35 GB	55%	0.0264	615	0
Qwen3.8-27B-NVFP4	`nvfp4-pack-quantized`	19.29 GB	35%	0.0950	615	0
Qwen3.8-27B-int4 ← this one	`pack-quantized`	18.31 GB	33%	0.1178	588	27

Mean relative error is ||dequant(W) - W|| / ||W||, averaged over a sample of quantized Linear layers, measured against the original BF16 weights. Lower is better.

This variant


Format	`pack-quantized`
Weight bits	4
Group size	128
Strategy	`group`
Linears quantized	588
Left in BF16	27
Shards	5
On disk	18.31 GB
Mean relative error	0.1178
Shape/dtype conformance failures	0

Layers left in BF16

27 Linear layers could not be quantized at group size 128: their in_features is not a multiple of 128. They are stored as BF16 and listed in the ignore list of config.json, so servers load them unquantized rather than misreading them.

They are all in the vision tower, so the language model is fully quantized:

model.visual.blocks.0.mlp.linear_fc2
model.visual.blocks.1.mlp.linear_fc2
model.visual.blocks.10.mlp.linear_fc2
model.visual.blocks.11.mlp.linear_fc2
model.visual.blocks.12.mlp.linear_fc2
model.visual.blocks.13.mlp.linear_fc2
... and 21 more

Use with vLLM

vllm serve dudeman2512/Qwen3.8-27B-int4

How this was made

Every produced tensor is checked for shape/dtype conformance against what the server expects, then reconstruction error is measured against the source BF16 weights, before anything is published. The numbers in the table above are those measurements — not estimates.