Ideogram releases its first open-weight image model
The 9.3B-parameter Ideogram 4 model was trained from scratch and adds a structured JSON prompting interface for text, layout, color and 2K image control.
By Ryan Merket ·
Why it matters
Ideogram is pushing against closed image systems by making weights available, a move that could appeal to developers and enterprises that need private customization rather than a hosted-only design tool.
Ideogram (@ideogram_ai) released Ideogram 4, its first open-weight text-to-image model, and published inference code and weights in a GitHub repository. The company says the 9.3B-parameter model is a foundation model trained from scratch, not a fine-tune or distillation of an existing checkpoint, and can be tried online at ideogram.ai.
https://x.com/ideogram_ai/status/2062202208700313872
The release is aimed at developers and design teams that want more control than a plain text prompt usually provides. Ideogram 4 was trained on structured JSON captions and introduces a JSON prompting interface for composition, typography, lighting, style, color palettes and spatial layout. The model supports explicit bounding-box coordinates, hex color palette conditioning and native images up to 2K resolution, with flexible aspect ratios from 256 to 2048 pixels on each side.
Ideogram says the model's central advantage is text and layout fidelity. Its GitHub documentation describes a flow-matching text-to-image model built on a single-stream Diffusion Transformer architecture, using Qwen3-VL-8B-Instruct as its text encoder and processing text and image tokens in one unified sequence. Plain text prompts still work, but the company says the best results come from the structured caption format or from its "magic prompt" system, which expands casual prompts into JSON before generation.
The competitive claim is that Ideogram 4 is the top open-weight image model across the company's cited evaluations. Ideogram says it leads open-weight models on Design Arena, ranked highest in a ContraLabs blind typography evaluation by professional designers, and is the top-ranked open-weight lab on LMArena's text-to-image leaderboard. Those rankings should be read as Ideogram's launch framing unless readers inspect the underlying third-party boards and benchmark methodology directly.
The weights are gated on Hugging Face under an Ideogram 4 Non-Commercial license, with nf4 and fp8 quantized variants listed in the model zoo. The nf4 version supports CUDA and Diffusers, while the fp8 version is listed for broader hardware support without Diffusers support. Ideogram says it plans to support more quantizations in the future, and also named distribution partners including Hugging Face (@huggingface), Cloudflare (@Cloudflare), Krea (@krea_ai) and Leonardo.Ai (@LeonardoAi).
