Ideogram does publish downloadable weights, and two conditions decide everything. The public release is quantized rather than full precision, and it is licensed for non-commercial use. This page covers what is on Hugging Face, what each licence permits, how a self-hosted stack splits into layers, and the documented commands.
What Ideogram has actually published
On 3 June 2026, Ideogram published inference code and weights together. Its framing is explicit: the weights are released so the research community can push visual intelligence forward. The GitHub README calls Ideogram 4 the company’s first open-weight text-to-image model, trained from scratch rather than fine-tuned.
What is on Hugging Face
The repository’s model zoo lists exactly two entries, both gated behind a licence acceptance.
- Ideogram 4 (nf4) — 9.3B parameters, nf4 quantization, supported hardware CUDA, Diffusers support yes.
- Ideogram 4 (fp8) — 9.3B parameters, fp8 quantization, supported hardware all, Diffusers support no.
Both carry the same licence: Ideogram 4 Non-Commercial. The README adds that more quantizations are planned.
What the vendor has not published
No weight file size appears in first-party material, so figures quoted elsewhere belong to whoever published them. No VRAM or GPU minimum is given: the model zoo names hardware categories, not numbers. No latency figure is published. And docs.ideogram.ai has no self-hosting page — that path returns the vendor’s own not-found page.
The three licences and who each fits
The licensing page sells three paths, differing in model access.
| Path | Model access | Self-hosting | Monthly volume |
|---|---|---|---|
| Non-commercial | Public, quantized | Non-commercial only | Research only |
| Self-serve commercial | Public, quantized | Yes | 10K–100K images |
| Enterprise | Full-precision | Yes | 100K+ or custom |
Self-serve is listed at $300 per month on annual billing with 10,000 images, excluding full-precision weights, hosted API access, implementation support and custom legal terms. Enterprise is the only tier with full precision.
What the agreement lets you do
The weights travel under the Ideogram Non-Commercial Model Agreement, updated 3 June 2026. Its grant is narrow: you may exercise the rights only for non-commercial purposes. That still covers a for-profit company testing in a non-production environment. Outputs are separate — Ideogram claims no rights in them.
Fine-tuning and modification
Fine-tuning, LoRAs and modifications are permitted for non-commercial use, and the licensing page says so. A derivative stays yours in its incremental rights, but redistributing it means passing on the same restrictions.
Redistribution and use limits
Redistribution requires no-less-restrictive terms, a copy of the agreement travelling with the model, the notice file keeping its attribution, and modified files flagged. You must not imply your derivative is an official Ideogram product. Banned uses include military and surveillance work, biometric processing, and high-risk decisions in finance, employment or healthcare. Outputs may not train a competing model.
The self-hosted deployment stack
Treat a self-hosted deployment as three layers, each with a different owner and licence.
Layer one: the weights
The weights are official, quantized and gated. Accept the gate on the model page, then authenticate, or the download fails with a 404. This layer carries the licence that changes if you buy self-serve rights.
Layer two: the runtime
The inference code is Apache 2.0 on GitHub, so the runtime is permissive although the weights are not. The pipeline exposes height, width, num_steps, guidance_scale, guidance_schedule, mu, std and seed; presets live in ideogram4.sampler_configs.PRESETS. Resolutions run from 256 to 2048 in multiples of 16, up to 6:1.
Layer three: the API layer
Exposing the model to others leaves the non-commercial grant behind. The agreement counts making the model available as a hosted service, by API or any other remote means, as distribution; customer-facing endpoints route to Enterprise. For the hosted path, the Ideogram API walkthrough documents the endpoints, and the JSON prompt guide covers the caption schema the CLI expects.
Running it: the documented commands
Setup is two steps: install, then authenticate so the gated weights download.
git clone https://github.com/ideogram-oss/ideogram4
cd ideogram4
pip install . # or: pip install -e . to edit src/ideogram4/
hf auth login # the weights are gated, so authenticate before the first run
export HF_TOKEN="hf_..." # alternative to hf auth loginThe CLI path
run_inference.py expands a plain prompt into the structured JSON caption the model expects. That expansion uses Ideogram’s hosted magic prompt API by default, which the README calls free, so a key is required.
python run_inference.py \
--prompt "a ginger cat wearing a tiny wizard hat reading a spellbook" \
--output out.png \
--quantization "nf4" \
--magic-prompt-key "$IDEOGRAM_API_KEY"For the highest quality, the README specifies a larger canvas and a named sampler preset; safety screening runs through Hive with two keys.
python run_inference.py \
--prompt "an isometric illustration of a tiny city floating in the clouds" \
--output out.png \
--quantization "nf4" \
--height 2048 --width 2048 \
--sampler-preset V4_QUALITY_48 \
--magic-prompt-key "$IDEOGRAM_API_KEY" \
--hive-text-key "$HIVE_TEXT_MODERATION_KEY" \
--hive-visual-key "$HIVE_VISUAL_MODERATION_KEY"The hosted API path
The same models are reachable over HTTP with no download and no hardware — the useful comparison when the licence blocks self-hosting.
API="https://api.ideogram.ai/v1/ideogram-v4/generate"
curl -s "${API}" \
-H "Api-Key: ${IDEOGRAM_API_KEY}" \
-F 'text_prompt=A poster on a wall with text that reads: "Everything you can imagine is real."' \
-F "resolution=1024x1024" \
-o generated.json
jq -r '.data[0].is_image_safe' generated.json
curl -sL -o poster.png "$(jq -r '.data[0].url' generated.json)"Choosing between weights and the API
Three axes decide it: licence, data residency and unit economics. Self-hosting keeps prompts and outputs inside your stack, the case the licensing page makes for data residency and latency; the hosted API is metered per image and needs no hardware. Non-commercial research points at the weights, commercial self-hosting at self-serve, customer-facing work at Enterprise. If you only need scheduled generation, webhooks may remove the reason to self-host; the tutorial covers the browser workflow.
Frequently asked questions
Does Ideogram publish model weights? Yes. Two quantized builds, nf4 and fp8, both at 9.3B parameters, on Hugging Face under the Ideogram Non-Commercial Model Agreement. Full precision is an Enterprise feature.
Can I use the weights commercially? Not under the open release. Self-serve adds commercial rights for those same public, quantized weights at $300 per month on annual billing.
How much VRAM do I need? The vendor publishes no figure. The model zoo lists hardware categories only.
Is the inference code restricted too? No. The licensing page lists Apache 2.0 inference code on GitHub, separate from the weights licence.