ollama run ssfdre38/gemma4-nano:e2b   # 3.1 GB
ollama run ssfdre38/gemma4-nano:e4b   # 4.7 GB (latest)
ollama run ssfdre38/gemma4-nano:26b   # 12 GB
ollama run ssfdre38/gemma4-nano:31b   # 13 GB

Or download GGUFs directly:

gemma4-e2b-q3ks-nano.gguf (2.9 GB) - Fits 4GB RAM devices
gemma4-e4b-q3ks-nano.gguf (4.3 GB) - Recommended default
gemma4-26b-q3ks-nano.gguf (11.4 GB) - High capability
gemma4-31b-q3ks-nano.gguf (12.8 GB) - 31B params, smaller than stock e4b!

Use with Ollama (from GGUF)

bash

# Download GGUF
wget https://huggingface.co/ssfdre38/gemma4-nano-gguf/resolve/main/gemma4-e2b-q3ks-nano.gguf

# Create Modelfile
echo "FROM ./gemma4-e2b-q3ks-nano.gguf
PARAMETER num_ctx 16384" > Modelfile

# Create model
ollama create my-nano -f Modelfile
ollama run my-nano

Specifications

Quantization: Q3_K_S (3.41 bits per weight)
Context: 16,384 tokens
Format: GGUF v3
Modality: Text-only (no vision encoder)

Performance

Nano models achieve 13% faster inference than gemma4-turbo on CPU with 8 threads.

Model	Size	Reduction vs Stock	Notes
E2b nano	3.1 GB	-57% (was 7.2 GB)	Fits 4GB RAM devices
E4b nano	4.7 GB	-51% (was 9.6 GB)	Recommended default
26b nano	12 GB	-29% (was 17 GB)	High capability
31b nano	13 GB	-32% (was 19 GB)	31B params < stock e4b size!

Documentation

Related Models

ssfdre38/gemma4-turbo - IQ4_XS with vision encoder

License

Apache 2.0 (same as Gemma 4 base models)

Built for the Gemma 4 Good Hackathon 2026 🦞