Qwen3-VL-30B-A3B-Instruct — wanda, 50% unstructured

Created with GBLM-Pruner using 128 C4 training windows of 2048 tokens (seed 0). Method: wanda. Measured decoder-matrix sparsity: 50.00000000%. The vision encoder, MoE routers, embeddings, normalization and output head remain dense. See pruning_report.json for the exact scope, counts and environment. GBLM uses the repository's additive activation-plus-absolute-gradient metric (summed L1 gradients scaled by 100). This is zero-weight pruning: tensor shapes and dense file sizes are unchanged. RealWorldQA evaluation is run separately; no accuracy claim is made until it completes.

from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration
path = "/path/to/this/model"  # or your uploaded Hugging Face repository
processor = AutoProcessor.from_pretrained(path)
model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    path, dtype="auto", device_map="auto", attn_implementation="sdpa")

The weights use native Transformers tensor names and require no custom modeling files. Prepared with Transformers 4.57.6. The base model is licensed under Apache 2.0.