Ideogram released Ideogram 4.0 on June 3, 2026, a 9.3-billion-parameter text-to-image model published with open weights on Hugging Face and inference code on GitHub. The release positions openness as a competitive strategy in an increasingly crowded generative-media market, and it arrives as the highest-ranked open-weight image model on the DesignArena leaderboard.
Key Highlights
- 9.3B parameters, open weights — model weights are published on Hugging Face and inference code lives in the ideogram-oss/ideogram4 repository.
- Native 2K resolution — supports any resolution from 256 to 2048 pixels (in multiples of 16) and aspect ratios up to 6:1, removing the need for a separate upscaling pass.
- Best-in-class text rendering — delivers the strongest in-image typography of any benchmarked open-weight release, ahead of much larger models.
- Structured layout control — bounding-box element placement, JSON prompting, color-palette conditioning via hex values, and multilingual text generation.
- Split licensing — code is released under Apache 2.0, while the weights are governed by an Ideogram Non-Commercial Model Agreement.
Details
Ideogram 4.0 ships in two quantized variants: an nf4 build that runs on CUDA-capable GPUs and supports Hugging Face Diffusers, and an fp8 build that works on a broader range of hardware. The model handles any resolution between 256 and 2048 pixels with automatic noise-schedule adjustment, and its inference script expands plain-text prompts into structured JSON through a "magic prompt" language model before generation begins.
That structured approach is central to the model's design focus. Instead of treating a prompt as a single block of text, Ideogram 4.0 accepts bounding-box specifications for where elements should sit, hex color values to condition the palette, and explicit instructions for in-image text — the kinds of controls that matter for logos, posters, and product mockups rather than generic illustration.
On the DesignArena leaderboard, Ideogram 4.0 ranks first among all open-weight models and ninth overall in the text-to-image arena, trailing only closed models from OpenAI and Google. At 9.3 billion parameters it delivers better text rendering than considerably larger open releases, including Qwen-Image at 20 billion parameters, FLUX.2 dev at 32 billion, and HunyuanImage 3.0 at 80 billion in a mixture-of-experts configuration.
Impact
The licensing split is the part worth reading closely. Releasing the inference code under Apache 2.0 while gating the weights behind a non-commercial agreement gives researchers, hobbyists, and teams evaluating in non-production environments genuine hands-on access, while keeping commercial deployment on a separate, paid track. For studios and agencies, that means the model can be tested locally and integrated into prototypes before any licensing commitment.
For builders, the structured JSON interface and editable-output direction matter more than raw image quality. Ideogram has signalled that an upcoming 4.0 release will return alpha channels and editable text layers directly from inference — no second pass, no masking step — so the model's output becomes a file production teams can hand off directly rather than a flat render to clean up later.
API pricing remains tiered for those who prefer the hosted route: roughly $0.03 per image on Turbo, $0.06 on the default tier, and $0.10 on Quality.
Background
Ideogram was founded in 2022 by former Google Brain researchers, including Mohammad Norouzi, and raised $80 million in early 2024. It shipped its first public model in August 2023 and built an early reputation specifically on text rendering — a weakness in most diffusion models of that era. Version 4.0 is the company's first release to put open weights at the center of that reputation, and it is already distributed across more than a dozen partner platforms including ComfyUI, Leonardo AI, and Picsart.
What's Next
The immediate question is adoption: open weights plus a strong text-rendering benchmark make Ideogram 4.0 an attractive base for fine-tuning and self-hosting, but the non-commercial weight license keeps production use behind a paywall. The promised editable-layer update — alpha channels and live text directly from inference — would push the model further toward being a design tool rather than an image generator, and that capability is what design teams will be watching for next.
Source: Ideogram