NF4 version with 24GB of VRAM can run it; individual players finally don't have to just watch the excitement.

Ideogram 4.0 is the first open-source weighted foundational image model, with 9.3 billion parameters. It uses a single-stream diffusion transformer architecture, with the text encoder being qwen3-vl-8b-instruct. It provides NF4 (CUDA support for 24GB of VRAM) and FP8 versions. The inference code is licensed under Apache 2.0; it is free for non-commercial/academic research, while commercial deployment requires a commercial license. The core innovation is a structured JSON prompt interface, enabling users to precisely control image layout, style, and composition through JSON strings. Benchmarks: 7bench 0.69 MIoU, X-Omni 0.97, and it ranks first among open-source models in preference blind testing.