Ideogram 4 withholds BF16 weights from public, ships FP8 only

Ideogram 4 ships with FP8 weights to the public while reserving BF16 weights for select partners, CEO Mohammad Norouzi confirmed during a June 21 YouTube stream with the ComfyUI team at the 37:45 mark. The move breaks from prior open-weight releases: Black Forest Labs released BF16 weights for FLUX.2 Dev despite its 60+ GB footprint, and Stability AI and other projects have historically shipped full-precision checkpoints alongside quantized variants.

The FP8-only public release means LoRA training and fine-tuning will operate at lower precision than the base model was trained in, typically degrading output quality and limiting what community fine-tunes can recover. Ideogram has not announced a timeline for wider BF16 availability or clarified partnership criteria. As of June 21, the HuggingFace model card lists only the FP8 checkpoint.