Ideogram · 2026-06-03 · major
Ideogram 4.0 — 9.3B Open-Weight Text-to-Image Foundation Model Lands With Apache-2.0 Inference Code, Native 2K Resolution, and Structured JSON Prompting That Tops DesignArena's Open-Weight Leaderboard
Ideogram released a 9.3B-parameter open-weight text-to-image foundation model trained from scratch with Apache-2.0 inference code, native 2K rendering, and structured JSON layout prompting on June 3.
Ideogram's first open-weight image model — 9.3B params, native 2K output, JSON-controlled layouts, and SOTA text rendering for design.
Key specs
| Parameters | 9.3B |
|---|---|
| GitHub stars | 886 |
| Design arena open weight rank | #1 |
| Design arena overall rank | #9 |
| Max resolution | 2048px (native 2K) |
| Min resolution | 256px |
| Hugging face likes | 193 (fp8) + 200+ (nf4) |
What is it?
Ideogram 4.0 is the San Francisco startup's first text-to-image model released with downloadable weights. It is a 9.3B-parameter flow-matching diffusion transformer trained from scratch, designed to render legible in-image text and complex layouts for design work. Open-source inference code ships under Apache 2.0 while the weights are gated behind a non-commercial license, with commercial use kept on Ideogram's paid API.
How does it work?
The model is a single-stream DiT that uses Qwen3-VL-8B-Instruct as its vision-language text encoder. It supports any resolution from 256 to 2048 pixels in multiples of 16, aspect ratios up to 6:1, and a structured JSON prompting interface that lets users pass bounding-box element placements and hex-coded color palettes alongside natural-language prompts. Two quantized checkpoints ship at launch — an NF4 build for CUDA + Diffusers and an FP8 build for any hardware.
Why does it matter?
Open-weight image models keep trailing closed APIs on legible text and design fidelity. Ideogram 4.0 ranks first among open-weight entries on DesignArena and ninth overall, and Ideogram says it beats Qwen-Image (20B), FLUX.2 dev (32B), and HunyuanImage 3.0 (80B) on text rendering at roughly a third of the parameter count. That puts a commercially restricted but downloadable SOTA in the hands of designers and researchers willing to swap the API for a local GPU.
Who is it for?
Designers, ML researchers, and image-tooling builders who want to run a near-SOTA text-render model on their own hardware
Try it
git clone https://github.com/ideogram-oss/ideogram4