StorySoftware
Ideogram releases 4.0, a 9.3B-parameter image model with open weights
Ideogram published the weights and inference code of Ideogram 4.0, a 9.3-billion-parameter text-to-image model trained from scratch. It was trained on structured captions that specify layout, colours and text, aiming at design work. Code is Apache 2.0, but the weights carry a non-commercial licence.
- Single-stream diffusion transformer with a Qwen3-VL-8B text encoder.
- Company claims 0.97 English OCR accuracy and the best text rendering among open-weight models.
- Second overall in Ideogram's designer arena, behind only GPT Image 2 medium.
- nf4 and fp8 versions published on Hugging Face; commercial deployment needs a separate licence.