Ideogram published Ideogram 4 together with weight files made available under an open approach, marking the first time weights were released publicly.

Model capabilities

The model natively generates images at 2K resolution and demonstrates reliable text rendering across styles and languages in diverse scenes.

It supports frame control using a color palette and bounding-box highlighting to influence composition and focal placement within generated images.

Weights, formats and local use

Weights were published in FP8 and NFP4 variants, while FP16 availability appears unlikely according to the release notes.

The model comprises 9.3B parameters; the FP8 weight file occupies 9.3 GB, and the text encoder adds 8.8 GB.

These sizes imply that the full model and encoder can fit into systems with about 24 GB of VRAM, enabling responsive inference on appropriately provisioned hardware.

Prompt tooling and inference options

In demonstrations, users can apply a prompt enhancer hosted by Ideogram servers or run the enhancer locally using the Qwen model for on-device processing.

Developers report three inference presets for trade-offs between speed and visual fidelity: Turbo with 12 steps, Default with 20 steps, and Quality with 48 steps.

License and usage

The distributed weights are provided under a non-commercial license, restricting use cases to non-commercial scenarios unless otherwise licensed by Ideogram.