Ideogram 4.0: A 9.3B open-weight image model (ideogram.ai)

Ideogram 4.0 has been launched as a significant advancement in the field of text-to-image generation, featuring an impressive 9.3 billion parameters and an open-weight framework. This model distinguishes itself by utilizing a Qwen3-VL-8B-Instruct text encoder, which pulls hidden states from 13 intermediate layers, allowing for smoother integration of text and image tokens through its 34-layer Diffusion Transformer (DiT). Unlike its predecessors, Ideogram 4.0 is trained specifically on structured JSON captions that offer detailed descriptions, including styling elements and bounding boxes, thus enabling the model to produce highly precise images based on input specifications. The implications for the AI/ML community are notable. Ideogram 4.0 not only enhances layout control, spatial reasoning, and text rendering abilities but does so with remarkable efficiency relative to its parameter size. With independent tuning for its conditional and unconditional sampling branches, the model achieves high-quality image generation while maintaining adherence to prompts. This innovation allows greater flexibility in how designers can specify and refine outputs, setting a new standard for open-weight generative models, with Ideogram 4.0 recently ranking second overall among graphic designer preferences against a backdrop of both open and closed-source competitors.