This is a modified (quantized, text-only) version of google/gemma-3-4b-it, packaged by LibrePower as Q4_K_M (Q6_K output head) GGUF for CPU inference on IBM Power (POWER9 VSX, POWER10/11 MMA-accelerated). No GPU required.

Gemma license. This model is governed by the Gemma Terms of Use and the Gemma Prohibited Use Policy, which apply to you as a downstream user. The weights have been modified (quantized to Q4_K_M, text-only, endianness-converted). NOT Apache-2.0 — unlike the rest of the LibrePower catalog.

Run it

Ubuntu / Debian ppc64le:

sudo apt install librepower-llama
lp-llama-models get gemma-3-4b      # or wget the LE file below
lp-llama-completion -m Gemma-3-4B-it-Q4_K_M.gguf -p "Hola!" -n 64 -t $(nproc)

IBM AIX 7.3 (big-endian):

dnf install llama-aix
wget https://huggingface.co/librepowerai/Gemma-3-4B-it-Power/resolve/main/Gemma-3-4B-it-Q4_K_M-be.gguf
lp-llama-completion -m Gemma-3-4B-it-Q4_K_M-be.gguf -p "Hola!" -n 64 -t $(nproc)

Files

Gemma-3-4B-it-Q4_K_M.gguf — little-endian (Ubuntu/Linux ppc64le)
Gemma-3-4B-it-Q4_K_M-be.gguf — big-endian (IBM AIX)

Good for

Multilingual chat and assistance (140+ languages), summarization, general instruction following. Text-only build (the original is multimodal).

Credits

Base model: Google Gemma (Gemma Terms of Use). Quantization & Power packaging: LibrePower.