The NF4 variant runs on 24GB consumer GPUs.
martin_casado#165
Today we published a technical blog post about Ideogram 4.0 — our goal is to enable more innovation and creativity.
It's a 9.3B Diffusion Transformer trained from scratch, paired with a frozen 8B VLM as text encoder. The nf4 checkpoint runs on a 24GB consumer GPU.
Thread 🧵
Many users celebrated Ideogram releasing technical details on its 9.3B open-weight diffusion model because the engineering enables strong open-source image generation that runs efficiently on consumer GPUs.
Fun fact about Ideogram 4.0: we haven't even scaled the model yet.
We're hiring! My team and I are at @CVPR for 3 days. DM me, swing by the booth, or come to our happy hour tonight: https://luma.com/vd0688ib?tk=wrYFBY
Today we published a technical blog post about Ideogram 4.0 — our goal is to enable more innovation and creativity.
It's a 9.3B Diffusion Transformer trained from scratch, paired with a frozen 8B VLM as text encoder. The nf4 checkpoint runs on a 24GB consumer GPU.
Thread 🧵
We believe openness drives innovation.
Both fp8 and nf4 checkpoints are in our repo, with the nf4 variant fitting on a single 24 GB GPU.
Huggingface: http://huggingface.co/collections/ideogram-ai/ideogram-4 Github: http://github.com/ideogram-oss/ideogram-4 Blog: http://ideogram.ai/blog/ideogram-4.0
Ideogram 4.0 is trained on detailed structured JSON prompts, with captions describing every element in the image. Elements can also include their bounding-box coordinates.
Giving the diffusion model as much information as possible improves output quality and training speed.
Ideogram 4.0 has four key components:
(1) a frozen Qwen3-VL-8B text encoder (2) a 34-layer single-stream DiT (3) a flow-matching Euler sampler with asymmetric CFG (4) a frozen FLUX.2 VAE
We thank the open-source community and give back by releasing this with open weights.
Ideogram 4.0 is designed to take structured JSON as input rather than unstructured text.
To make that easy, we've built a prompt-enhancement model that turns natural language into our JSON format. It's free in our API, and we've open-sourced a system prompt so you can get the same conversion from off-the-shelf LLMs too.
As the text encoder we use Qwen3-VL-8B, a vision-language model — so the same encoder can also take images for editing, not just text.
Our DiT consumes hidden states from 13 intermediate layers concatenated along the feature dimension, instead of a single hidden state.
The result is the best open-weight image model available, closing the gap with closed-source foundation models.
This is a 9.3B model — far from scaling limits. We expect further gains as we scale.
The only trainable component of Ideogram 4.0 is a 9.3B-parameter single-stream DiT.
We use a single-stream DiT with self-attention, QK-RMSNorm, and 3D Multimodal RoPE that places text and image tokens in one positional frame, for a flow-matching objective.
@ideogram_ai Fun fact about Ideogram 4.0: we haven't even scaled the model yet 🔥🔥
@ideogram_ai
@hassan_mar3069 You need to provide JSON prompts
@ideogram_ai Its awesome. Wish it could take an image as prompt too.
@ideogram_ai I really love Ideogram 4.0. It is fresh air in open source models. I made these few minutes ago at home. Thank you for the release.
@ideogram_ai Amazing job! A great day for Open Source!
@mo_norouzi @ideogram_ai Congratulations @mo_norouzi and team! Fantastic to see this in the open
@Tnimbus coming soon!
@ComfyUI ❤️
@ideogram_ai Pro-level image AI now runnable on a 24GB GPU.
@ideogram_ai Running on a normal GPU makes it more accessible
@Tnimbus @SubarcticRec Currently you can only use image as init image, but we're working on image editing.