Ideogram drops open-weight Ideogram 4.0 image model

Ideogram 4.0 is Ideogram's first open-weight text-to-image foundation model, weighing in at 9.3B parameters. It lets you download the weights, run inference on your own hardware, fine-tune on custom datasets, and output design-ready images with dense text rendering, native 2K resolution, and background transparency.

Open-source image generators usually fall apart the moment you ask them to spell words or arrange items cleanly across a canvas. Ideogram 4.0 tackles this head-on with a structured JSON prompting interface, explicit bounding-box layout coordinates, and color palette controls.

What it does and benchmark results

The model is trained from scratch rather than fine-tuned on an older architecture. At 9.3B parameters, it delivers text rendering that beats much larger models on the X-Omni OCR benchmark, including the 20B Qwen-Image, the 32B FLUX.2 dev, and HunyuanImage 3.0 at 80B MoE.

Third-party leaderboards back up the claims. On DesignArena's image Elo board, Ideogram 4.0 ranks as the number one open-weight model, trailing only proprietary models from OpenAI and Google. On LMArena's text-to-image board, Ideogram stands as the top open-weight lab and lands in the top five overall.

In a blind typography test run by ContraLabs with ten top-earning freelance designers, Ideogram 4.0 took first place with a 47.9% win rate. That cleared Gemini 3.1 Flash Image Preview at 30.0%, FLUX.2 max at 15.5%, and Grok Imagine 1.0 at 15.0%. When designers were asked if they would use the output in real client work, Ideogram 4.0 led with a 3.55 out of 5 score.

On the 7Bench layout control benchmark, it even outperformed leading closed-source alternatives.

How to run it

Ideogram released two quantization builds: an nf4 version compatible with CUDA and Diffusers, and an fp8 version that runs across all supported hardware.

The inference code is hosted on GitHub, while the model weights sit behind a gated repository on Hugging Face. You will need to accept the license agreement on Hugging Face and log in with an access token before pulling the files.

To bridge human text and structured generation, the pipeline uses a magic-prompt LLM to turn plain text into the structured JSON caption the base model expects. By default, it hits Ideogram's free hosted API, but you can also route the prompt expansion to a local LLM using Ideogram's open-sourced system prompt.

For max quality, the script supports running native 2048×2048 generation. If you prefer zero setup, the model is already available on ideogram.ai across all tiers and through the official API.

Why it matters

If you build graphic assets, marketing banners, or UI mockups, Ideogram 4.0 is a massive upgrade over running older local checkpoints. Being able to fine-tune a model that natively respects text spelling and bounding-box coordinates unlocks real production workflows on local GPUs.

The catch is the license: weights are shared under the Ideogram 4 Non-Commercial license, meaning true commercial deployment still routes through their web platform or API. Still, showing that a 9.3B model can outperform 30B+ giants in text rendering fundamentally resets expectations for open-weight creative tools.