Expand description
Qwen3.5 vision encoder
Architecture: Conv3d patch embedding → 27 transformer blocks (2D RoPE on Q/K) → VisionPatchMerger (2×2 spatial merge) → out_hidden_size projection.
Output: [total_image_patches, out_hidden_size]
Qwen3.5 vision encoder
Architecture: Conv3d patch embedding → 27 transformer blocks (2D RoPE on Q/K) → VisionPatchMerger (2×2 spatial merge) → out_hidden_size projection.
Output: [total_image_patches, out_hidden_size]