Ideogram has released its first open-weight image model, combining structured JSON prompting with benchmark-leading typography and layout capabilities aimed at professional design workflows.
Ideogram has released Ideogram 4.0, its first open-weight foundation image model, introducing a 9.3-billion-parameter system built specifically for professional design work. Trained from scratch and available with public weights for local deployment, the model targets applications ranging from posters and packaging mockups to social media graphics and branded content.
A key differentiator is its exclusive use of structured JSON captions during training, with inference also relying on JSON-formatted prompts. This enables direct control over colour palettes, bounding-box-based layout positioning, and typed text rendering, allowing designers to specify visual elements with greater precision. Training and inference operate using the same structured format, helping improve consistency between prompts and generated outputs.
According to Ideogram, the model is the top-ranked open-weight system on Design Arena, trailing only proprietary GPT and Gemini image-generation systems overall. In a ContraLabs blind evaluation involving 10 professional designers, Ideogram 4.0 was selected as the best model 47.9% of the time, outperforming Nano Banana 2 at 30.0%, FLUX.2 Max at 15.5%, and Grok Imagine 1.0 at 15.0%. It also achieved a client-work usability score of 3.55 out of 5, compared with 2.84 for Nano Banana 2.
Ideogram claims the model delivers the strongest text-rendering performance among benchmarked open-weight models, surpassing larger rivals including Qwen Image, FLUX.2 Dev and HunyuanImage 3.0. It also demonstrated strong results in prompt alignment, spatial reasoning and layout-control benchmarks.
Available through Hugging Face in FP8 and NF4 quantisations, the NF4 version can run on a single 24GB GPU and supports fully offline workflows through Diffusers integration.
