Gemma 4 Turbo

A fully optimized Gemma 4 family built from bf16 source weights — smaller, faster, and fully multimodal out of the box.

Why Turbo?

Google’s stock Gemma 4 on Ollama ships with Q4_K_M quantization applied to already-quantized weights. Gemma 4 Turbo starts from the original bf16 source and applies IQ4_XS (4.25 bpw non-linear quantization), producing better quality at a smaller size. Full vision + thinking capabilities are preserved.

Quick Start

ollama run ssfdre38/gemma4-turbo          # e4b (recommended)
ollama run ssfdre38/gemma4-turbo:e2b      # 8GB RAM machines
ollama run ssfdre38/gemma4-turbo:26b      # high quality
ollama run ssfdre38/gemma4-turbo:31b      # maximum quality

Vision & Multimodal

All tags include the full vision encoder — text, images, and thinking are all supported:

ollama run ssfdre38/gemma4-turbo "describe this image" /path/to/image.jpg

Performance

Benchmarked on CPU (Intel Xeon E-2236, 6C/12T, no GPU). All runs clean with no competing processes.

Size & RAM (e4b)

Tokens per Second (e4b, 8 threads)

Flash attention delivers the biggest gain on short conversations — the most common real-world use case.

Prefill Speed (time to first token, 8 threads, no flash attn)

Speed Tip — Enable Flash Attention

Set this environment variable before starting Ollama for roughly 2x faster token generation on short and medium-length conversations:

Windows:

[System.Environment]::SetEnvironmentVariable("OLLAMA_FLASH_ATTENTION", "1", "Machine")

macOS/Linux:

echo 'OLLAMA_FLASH_ATTENTION=1' >> ~/.bashrc   # or ~/.zshrc

Then restart Ollama. Also recommended:

OLLAMA_KV_CACHE_TYPE=q8_0    # halves KV cache RAM usage

What Makes This Different From a Repackage

• IQ4_XS applied to bf16 source — not re-quantizing already-quantized weights. Source weights downloaded from bartowski’s bf16 GGUFs, quantized with llama.cpp llama-quantize.
• Vision encoder preserved — original Ollama blobs contain 1411 vision tensors (a.*, v.*, mm.*) plus gemma4.vision.* KV metadata. These are merged back into the IQ4_XS LM weights so nothing is lost.
• Tuned defaults — num_thread 8, num_batch 512, num_ctx 16384 benchmarked and optimized for CPU inference. Setting threads to the logical processor count (e.g. 12 on a 6C/12T CPU) kills eval speed via hyperthreading contention — physical core count is the sweet spot.

Technical Details

• Quantization: IQ4_XS (4.25 bpw, non-linear, importance matrix sampling)
• Source: bartowski bf16 GGUFs → llama.cpp b9050 llama-quantize
• Vision: merged from original gemma4:e4b/e2b/26b/31b Ollama blobs (2131 tensors total for e2b/e4b; 1076 for 26b/31b)
• Architecture: gemma4 with full multimodal projector

See It In Action — Ash Bot

Ash is a self-hosted Discord AI bot built on .NET 10 that ships with ssfdre38/gemma4-turbo as its default model. It’s the reference implementation for running this model in a real application.

What Ash does: - 💬 Natural conversation with a consistent personality in your Discord server - 🧠 Long-term memory across sessions (memories.json) - 🔧 20 built-in tools — web search, YouTube Music, file ops, code execution, reactions, DMs, and more - 🤖 Autonomous initiative — speaks unprompted on a configurable interval - 🦙 Fully local — no cloud AI APIs, everything runs through Ollama on your own machine

Quick start:

git clone https://github.com/ssfdre38/ash-bot
cd ash-bot
# Windows:
setup.bat
# Linux/macOS:
./setup.sh

Ash will auto-pull ssfdre38/gemma4-turbo on first launch if it isn’t already installed.

→ github.com/ssfdre38/ash-bot

License

Derived from google/gemma-4-e4b-it and family. Usage governed by the Gemma Terms of Use.
Quantization and optimization work by ssfdre38.