Ideogram Drops Ideogram 4, a 9.3B Open Model Beating 80B Rivals on Text
Ideogram releases its first open-weight model: a 9.3B single-stream DiT that beats much larger rivals on typography and design tasks.
Takeaways
- Ideogram releases its first open-weight text-to-image model, a 9.3B single-stream DiT with 34 layers.
- Uses Qwen3-VL-8B-Instruct as text encoder, pulling features from 13 intermediate layers.
- Trained exclusively on JSON captions with bounding boxes, color palettes, and per-element styling.
- Beats FLUX.2 (32B) and HunyuanImage 3.0 (80B MoE) on text rendering benchmarks.
- Ships in NF4 (24GB GPU) and FP8 quantizations under a non-commercial license.
- Code on GitHub, weights gated on Hugging Face, native resolution up to 2048px.
Ideogram just dropped its first open-weight text-to-image model, and it is aimed squarely at one thing that most diffusion releases still get wrong: designers actually being willing to ship the output. Ideogram 4 is a 9.3B parameter Diffusion Transformer trained from scratch, shipped with code on GitHub and weights on Hugging Face in two quantized flavors. The pitch is that it punches well above its parameter class on the things that matter for posters, logos, and layouts: text rendering, spatial control, and typography.
According to the technical writeup, this is not a fine-tune of FLUX or anything else floating around. Ideogram 4.0 is a 9.3B parameter open-weight text-to-image model where text and image tokens share the same projections at every layer of a 34-layer DiT. That single-stream design is the architectural backbone, and the team made two unusual choices around it.
A vision-language model doing the text encoding
The first surprise sits where you would normally find CLIP or T5. The text encoder is Qwen3-VL-8B-Instruct, a vision-language model, and the DiT consumes hidden states from 13 of its intermediate layers concatenated along the feature dimension, instead of a single hidden state or no external encoder at all.
The intuition is that a VLM already understands visual concepts in a way pure text encoders never quite do, and pulling features from many depths gives the diffusion model both the surface token-level information and the deeper compositional reasoning at once. It is also a near-free architectural lift now that strong open VLMs exist, which probably explains why this pattern is going to spread.
The second choice is the prompt format itself. The model is trained exclusively on structured JSON captions with per-element styling and optional bounding boxes and color palettes, and the reference inference pipeline parses every prompt as JSON and validates it against the schema before generation. Plain text still works, but a hosted "magic prompt" LLM rewrites it into JSON before the diffusion step runs.