What GPU do you need to run ZoomGain-9B?

9.41B parameters (Hugging Face, accessed 2026-11-11). Q4_K_M GGUF: 5.63 GB (source). Architecture config (Hugging Face config.json): 8 attention layers (the rest are recurrent) × 4 KV heads × 256 head dim, so about 32 KB of FP16 KV cache per token of context (how we estimate context).

inclusionAI vision-language model, RL-trained from Qwen3.5-9B on inclusionAI's ZoomGain data for fine-grained visual perception: details that only show when you zoom in, such as attributes, counting and OCR (paper: arXiv 2611.14076). Its config matches Qwen3.5-9B's except that the MTP layer is removed (mtp_num_hidden_layers 0), so there is no built-in MTP drafting. Image input needs the mmproj file, which the fit leaves out. Q4_K_M is inclusionAI's official GGUF (ZoomGain-9B-Q4_K_M.gguf, 5,629,108,832 bytes); every size is in the table below. License: the Hugging Face card and repo state none for these weights, so confirm the terms with inclusionAI before commercial use (details below). Release date is the Hugging Face repo creation date.