paralaif/Huihui-Qwen3-VL-8B-Instruct-abliterated-comfy-int8-convrot

Huihui Qwen3-VL-8B-Instruct Abliterated — Quantized checkpoints

Quantized variants of Aragduhl/Huihui-Qwen3-VL-8B-Instruct-abliterated-comfy, which packages the full vision tower and language model for ComfyUI. The base repository lists the model under the Apache-2.0 license.

These checkpoints were produced from the BF16 single-file model using Comfy-Org/comfy-model-tools and Comfy Kitchen.

Files

File	Format	Size	Quantized weights
Huihui-Qwen3-VL-8B-Instruct-abliterated_w4a8.safetensors	W4A8	6.46 GiB	252 language-model linear layers; embeddings in INT8
Huihui-Qwen3-VL-8B-Instruct-abliterated_int8_convrot.safetensors	INT8-ConvRot	9.29 GiB	252 language-model linear layers and embeddings in INT8

The vision tower and lm_head are retained in BF16 in both files. W4A8 uses 4-bit weights with INT8 activations at runtime. INT8-ConvRot stores the language-model linear weights in INT8 with ConvRot; activations are quantized at runtime.

Quantization measurements

These are relative weight-reconstruction errors reported by the exporter, not output-quality scores:

• W4A8: 7.313% mean across 252 quantized layers.
• INT8-ConvRot: 0.888% mean and 1.021% maximum across 252 quantized layers.

The safetensors headers and quantization markers were checked. No ComfyUI inference or output-quality benchmark was run.