Mixed low-bit GGUF of google/gemma-4-26B-A4B-it, used by the BakAI Android app, which downloads it on first launch.
| tensors | type |
|---|---|
| routed experts gate/up | IQ2_XXS (2.06 bpw) |
| routed experts down | IQ4_NL (llama.cpp needs 256-divisible columns; Gemma 4's down projection has 704) |
| attention, token embedding | Q4_K |
| dense MLP | Q4_K / Q5_0 |
Average 3.05 bpw, 9.63 GB, split into 7 parts with llama-gguf-split (load the first part;
llama.cpp finds the rest). Importance matrix: bartowski's for the same model.
Quality against the original has not been measured yet.
License: Apache 2.0, same as the base model (Gemma 4 license).
