Is Ideogram 4.0 open source? Apache code, non-commercial weights

Partly. Ideogram 4.0 is an open-weights model, not an open-source one in the usual sense. Ideogram's licensing page says the inference code is Apache 2.0 and published on GitHub, but the model weights you download from Hugging Face are free only under the Ideogram Non-Commercial Model Agreement, for research, evaluation and personal projects.

If you want to self-host the weights and sell what they make, you need one of Ideogram's paid commercial licences. Sume's Image API does not run your own weights and does not list Ideogram 4.0 at all: its catalog lists ideogram/ideogram-v3. Both facts were read from vendor pages and Sume's docs on 2026-10-03.

What does Ideogram's own licensing page say?

The licensing page splits access into three tiers. The wording below is the page's, trimmed to the points that decide whether you can ship work.

What is actually on Hugging Face?

The Ideogram organisation page lists three Ideogram 4 repositories dated June 3 and 4: an nf4 build, an fp8 build, and an nf4 build for Diffusers. The fp8 model card names the licence as Ideogram 4 Non-Commercial and points to a LICENSE.md file; the same card says you must accept a gate before the files appear.

• The model has 9.3B parameters and a single-stream diffusion transformer of 34 layers, according to the card.
• Its text encoder is Qwen3-VL-8B-Instruct, not CLIP.
• It was trained exclusively on structured JSON captions; plain-text prompts work, but the card says JSON gives the best results.
• Resolution runs from 256 to 2048 pixels a side in multiples of 16, with aspect ratios up to 6:1.

Why does "open" mislead here?

Three different things get called open: published code, downloadable weights, and permission to use the output commercially. Ideogram 4 has the first two. The third is a separate paid grant, and the licensing page prices it per month and by monthly image volume. Anyone who reads "Apache 2.0" in a headline and assumes the weights are Apache 2.0 will get the licence wrong.

The consequence is practical. A designer prototyping an ad concept at home is inside the free licence. An agency generating client deliverables from a self-hosted copy is not, unless it buys the self-serve or enterprise tier. This is a reading of vendor pages, not legal advice; the LICENSE.md in the repository governs.

What does this change if you generate on Sume?

Sume's Image API is a hosted catalog. You pick a model id from GET /v1/images/models, send a prompt, and pay a per-image price that already includes Sume's margin. There is no field for uploading weights, so nothing about Ideogram's weights licence applies to a Sume call. The licence question moves to whatever model you choose from the catalog.

What you give up is the specific model: the catalog row is Ideogram V3, a different release. If the point is text rendering, compare V3 against another catalog model on your own prompt before assuming it matches 4.0, as in the post on Ideogram 4.0 on Sume. This script prints the Ideogram ids Sume lists right now:

import os
import requests

r = requests.get(
    "https://api.sume.com/v1/images/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
ids = [m["id"] for m in r.json()["data"]]
print([i for i in ids if "ideogram" in i])
print(len(ids), "image models listed")

Which route should you take?

Use the weights for research and private experiments, buy a commercial tier if you will self-host at volume, and use a hosted catalog model when you need an image today without managing a GPU. If you pick the paid licence, the next question is cost: the per-image price of Ideogram V3 on Sume gives you the number to set against the monthly fee.