convert_hf_to_gguf.py exports the language decoder only. On this model it wrote 851 of 1185 tensors; the remaining 334 are the vision tower and live in mmproj-q27b_vl-f16.gguf. Nothing in the converter's output says so — it reports Model successfully exported and exits zero. Load the mmproj alongside the main GGUF or the model is blind.

None of these fit a 12 GiB GPU with room for KV cache; Q3_K_M is the closest and still exceeds it. A 27B on 12 GiB needs IQ3 or lower, which is untested here — low-bit quantisation of the Gated DeltaNet layers is exactly where this architecture has already surprised us once (see the parent card on GPTQ).

Receipts, method and caveats: see the bf16 parent.