PsyGuard-Gemma-4-26B-A4B-alpha

Experimental Gemma-4-26B-A4B fine-tune (MoE, ~4B active). Built on unsloth/gemma-4-26B-A4B-it.

alpha — experimental. Internal evals only. Dogfood at your own risk.

Format: Q5_K_M GGUF (llama.cpp / LM Studio / Ollama), text.

llama-server -m PsyGuard-Gemma-4-26B-A4B-alpha-Q5_K_M.gguf -ngl 99 --jinja --reasoning off

(Serve with --reasoning off for direct answers.) Requires a recent, Gemma-4-aware build of llama.cpp / LM Studio / Ollama.