As of March the 18th 2026 the early version of the V8 model is now being tested with the community. It follows prompts better, runs about 5x faster, supports native 2K output, and improves text rendering and personalization.
The model handles detailed instructions well and can still give unexpected results at times. It learns user style through personalization, style references and moodboards, leading to more consistent and detailed images. Text inside quotes works more reliably, and image generation is much faster.
The web interface was updated to match the speed. It now includes a smoother chat flow, a Grid Mode for viewing larger image sets, and sidebar settings so users can adjust things without blocking the screen. More updates are planned.
V8 supports different aspect ratios along with options like --chaos, --weird, --exp and --raw. It also works with older V7 personalization profiles, moodboards and style references.
A new --hd mode allows native 2K images, and --q 4 adds extra consistency. Some modes like HD, Q4, style references and moodboards take about 4x longer and cost more for now. Relax mode is not available yet, but work is ongoing to bring it back with cheaper options.
Users can help improve the model by rating images. In the create view, they can select an image and choose like or dislike, or use hotkeys for faster rating.
For more simple or controlled results, it is suggested to use --raw or rely on moodboards and style references. The default look is still being refined, so users may get better results by pushing stylization higher and using longer prompts.
If you'd like to access this model, you can explore the following possibilities:
