Ollama has added experimental image generation to its local inference toolkit, extending the popular LLM runner into text-to-image territory. The feature launched January 20 with macOS support only.
Two models ship at launch. Z-Image Turbo, a 6 billion parameter model from Alibaba's Tongyi Lab, targets photorealistic output and handles bilingual text rendering in English and Chinese. FLUX.2 Klein from Black Forest Labs comes in 4B and 9B variants, optimized for speed and readable typography in generated images. The 4B version runs under Apache 2.0; the 9B falls under a non-commercial license.
The implementation leans into Ollama's terminal-native design philosophy. Images save to the current working directory, and terminals with inline image support can render results without leaving the shell. Configuration happens through CLI commands: /set width, /set height, and step count adjustments for balancing quality against generation time. Seeds enable reproducible outputs.
Windows and Linux support is listed as coming soon. The blog post also promises additional models and image editing capabilities in future updates.
The Bottom Line: Ollama's move into image generation brings text-to-image workflows into the same local-first environment developers already use for LLMs.
QUICK FACTS
- Z-Image Turbo: 6B parameters, Apache 2.0 license
- FLUX.2 Klein: 4B (Apache 2.0) and 9B (non-commercial) variants
- macOS only at launch; Windows/Linux pending
- Terminal preview in Ghostty and iTerm2
- Negative prompts and seed control available
