> run with ollama
Ollama is the fastest way to run Gemma 4 locally: one ollama run gemma4 pulls the weights and starts a chat, with a REST API on port 11434 for apps. Tags map to sizes (gemma4:e2b through gemma4:31b, plus -mlx builds for Apple Silicon), and the bare gemma4 alias is E4B. Models live under ~/.ollama/models.
// run with ollama
6 commands$ ollama pull gemma4:31b$ ollama run gemma4$ ollama run gemma4 "describe ./chart.png"$ ollama run gemma4:e4b-mlx$ curl http://localhost:11434/api/generate -d '{"model":"gemma4","prompt":"roses are red","stream":false}'$ printf 'FROM gemma4\nSYSTEM "You are a terse coding assistant."\n' > Modelfile// faq
Which model does ollama run gemma4 download?
The bare gemma4 tag resolves to the E4B build (about 9.6 GB), which Ollama marks as latest. Ask for gemma4:e2b, gemma4:12b, gemma4:26b, or gemma4:31b explicitly to get a different size. gemma4:26b is the A4B Mixture-of-Experts, not a dense 26B.
How do I send Gemma 4 an image with Ollama?
Include the image's file path inside the prompt string, for example ollama run gemma4 "describe ./chart.png". Gemma 4 is natively multimodal, so all sizes read images; E2B, E4B, and 12B also accept audio. Put the image reference before your instruction for the best results.
Does Gemma 4 on Ollama expose an OpenAI-compatible API?
Yes. Alongside the native /api/generate and /api/chat endpoints on port 11434, Ollama serves an OpenAI-compatible surface at http://localhost:11434/v1/chat/completions. Point any OpenAI client at that base URL with model set to gemma4 and any dummy API key.
Where does Ollama store the downloaded Gemma 4 models?
Under ~/.ollama/models as content-addressed blobs plus a manifest per tag. Set the OLLAMA_MODELS environment variable to move them to a bigger disk. Use ollama list to see what is installed and ollama rm gemma4:31b to free space.