Gemma 4 Turbo is a purpose-built quantization of Google's Gemma 4 (9B) model. Built from original bf16 source weights using IQ4_XS non-linear quantization (4.25 bpw) with full multimodal/vision capability preserved. 36% smaller than stock Ollama builds (6.1 GB vs 9.6 GB). Runs on CPU-only hardware — no GPU required. Part of the ash-server ecosystem (github.com/ssfdre38/ash-server). Submitted to the Kaggle Gemma 4 Good Hackathon 2026.