RFD 1155: Gemma 4 reaches the device as towers, not as a decoder
Problem
Gemma 4 is wanted for language, vision and sound in one model. At 12 B it is 6.60 GB at four bits and does not fit at any other precision, so it inherits RFD 1128’s open question.
Two things stop it that memory does not see. The checked-out artifact is GGUF, which carries no graph, so nothing converts. And autoregressive decode grows its sequence and cache every token while a dataflow part compiles fixed shapes, which is the obstacle RFD 1126 names: control flow rather than precision.
Decision
Abandoned on 2026-08-28. The accelerator work is scoped to rf-detr keypoint and RFD 1157. Gemma 4 ranked last at 4 of 25 against that model’s 18, scoring zero on shape, reference and clear at once, and the gap is a reference implementation. 3-interactor/llama-cpp-npu-vision-upstream compiles an encoder for Qwen3-VL and nothing else, and RFD 1157’s model sits on that family. One encoder serves both, where Gemma 4 needs its own with no precedent. Qwen3-VL-4B is also comfortable at eight bits, so it never meets the four-bit question.
What Qwen3-VL does not carry is sound. If audio is a requirement it does not replace Gemma 4, and no Hailo reference for an audio model is known here. If Gemma 4 returns, plan the towers and not the decoder.