Downloads · 30 days
225
4% of all-time downloads
gemma4-turbo-gguf
ssfdre38/gemma4-turbo-gguf
gemma4-turbo-gguf is a machine learning model from ssfdre38. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Optimized Gemma 4 models with 51% faster CPU inference through IQ4XS quantization and performance tuning.
All-time downloads
5.6K
Public
Repo size
50.6 GB
Likes
4
Public
Repo makeup
Click a slice to open those files.
.gguf50.6 GB · 100%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- May 12, 2026
- Updated
- Jun 5, 2026
- SHA
- f6adfb8d
Model card
From the Hugging Face model README
gemma4-turbo GGUF Files
Optimized Gemma 4 models with 51% faster CPU inference through IQ4_XS quantization and performance tuning.
Quick Start with Ollama
ollama run ssfdre38/gemma4-turbo:e2b # 4.3 GB
ollama run ssfdre38/gemma4-turbo:e4b # 6.1 GB (latest, recommended)
ollama run ssfdre38/gemma4-turbo:26b # 15 GB
ollama run ssfdre38/gemma4-turbo:31b # 18 GB
Or download GGUFs directly:
- gemma4-e2b-iq4xs-turbo.gguf (4.3 GB) - Smallest with vision
- gemma4-e4b-iq4xs-turbo.gguf (6.1 GB) - Recommended default
- gemma4-26b-iq4xs-turbo.gguf (15 GB) - High capability
- gemma4-31b-iq4xs-turbo.gguf (18 GB) - Maximum performance
Use with Ollama (from GGUF)
# Download GGUF
wget https://huggingface.co/ssfdre38/gemma4-turbo-gguf/resolve/main/gemma4-e4b-iq4xs-turbo.gguf
# Create Modelfile
echo "FROM ./gemma4-e4b-iq4xs-turbo.gguf
PARAMETER num_ctx 16384" > Modelfile
# Create model
ollama create my-turbo -f Modelfile
ollama run my-turbo
Specifications
- Quantization: IQ4_XS (4-bit with importance matrix)
- Context: 16,384 tokens (128K capable)
- Format: GGUF v3
- Modality: Text + Vision (multimodal)
- Audio: Supported
Performance
gemma4-turbo achieves 51% faster inference compared to stock Gemma 4 models on CPU (8 threads, Intel Xeon E-2236).
| Model | Size | Reduction vs Stock | Speed Improvement | Multimodal |
|---|---|---|---|---|
| E2b turbo | 4.3 GB | -40% (was 7.2 GB) | +51% | ✅ Vision+Audio |
| E4b turbo | 6.1 GB | -36% (was 9.6 GB) | +51% | ✅ Vision+Audio |
| 26b turbo | 15 GB | -17% (was 18 GB) | +51% | ✅ Vision+Audio |
| 31b turbo | 18 GB | -10% (was 20 GB) | +51% | ✅ Vision+Audio |
Key Features
✅ Full multimodal support - Vision and audio capabilities retained
✅ IQ4_XS quantization - Advanced 4-bit compression with quality preservation
✅ 51% faster inference - Optimized for CPU-only environments
✅ Windows-optimized - Tested extensively on Windows systems
✅ Tool calling - Built-in function calling support
Use Cases
- Local AI assistants with vision capabilities
- Chat applications requiring multimodal input
- Tool-calling workflows with image understanding
- Windows systems without GPU acceleration
- Privacy-focused deployments (fully offline capable)
Comparison: Turbo vs Nano
| Feature | Turbo | Nano |
|---|---|---|
| Quantization | IQ4_XS (4-bit) | Q3_K_S (3-bit) |
| Vision | ✅ Yes | ❌ No |
| Audio | ✅ Yes | ❌ No |
| Size (e2b) | 4.3 GB | 3.1 GB |
| Speed (e2b) | +51% vs stock | +13% vs turbo |
| Quality | Higher (4-bit) | Good (3-bit) |
| Use Case | Multimodal AI | Text-only, mobile |
Choose turbo for multimodal AI with vision/audio support.
Choose nano for ultra-lightweight text-only deployments on mobile/edge devices.
Documentation
Technical Details
Quantization Method
- IQ4_XS: 4-bit integer quantization with importance matrix
- Quality retention: 95%+ on most benchmarks
- Memory bandwidth: 60% reduction vs FP16
- Cache efficiency: Improved L2/L3 utilization
Testing Environment
- CPU: Intel Xeon E-2236 (6 cores @ 3.4GHz)
- RAM: 32 GB DDR4
- OS: Windows Server 2019
- Ollama: v0.5.1+
- Threads: 8 (6 cores + hyperthreading limited)
Performance Methodology
- Stock Gemma 4 baseline (FP16 or Q4_0)
- gemma4-turbo (IQ4_XS)
- Same prompt, 50 token generation
- Average of 10 runs
- Result: 51% faster tok/s on CPU
License
Apache 2.0 (same as Gemma 4 base models)
Citation
@software{gemma4-turbo,
author = {Daniel Elliott (ssfdre38)},
title = {gemma4-turbo: Optimized Gemma 4 Models with 51% Faster CPU Inference},
year = {2026},
url = {https://ollama.com/ssfdre38/gemma4-turbo}
}
Built for the Gemma 4 Good Hackathon 2026 🦞
Developer: ssfdre38 (Daniel Elliott)
Website: https://ssfdre38.xyz
- License
- apache-2.0
- Languages
- en
- Created
- May 12, 2026
- Updated
- Jun 5, 2026