ahmedabv/

qwen3.5-4b-base-blind-spots

	Qwen3.5-4B-Base Blind Spots Dataset




	
	Model Tested

Qwen/Qwen3.5-4B-Base — a 4B-parameter vision-language base (pretrained-only) model released in February 2026 by the Qwen team. Key specs:

Architecture: Hybrid Gated DeltaNet + sparse Mixture-of-Experts (32 layers, hidden dim 2560) Context length: 262,144 tokens natively, extensible to 1,010,000 Modality: Text + Vision (early fusion on multimodal tokens) Languages: 201 languages and dialects Type: Pre-trained base model —… See the full description on the dataset page: https://huggingface.co/datasets/ahmedabv/qwen3.5-4b-base-blind-spots.