# Ideogram 4 ships Apache-2.0 code, non-commercial weights, and a default prompt path that calls a hosted API

> A 9.3B text-to-image model released in June 2026 as a foundation model trained from scratch, with a structured JSON prompting interface and a benchmark section full of third-party leaderboard claims. Two things qualify the openness: the weights are gated and non-commercial, and the default prompt expansion happens server side at Ideogram's API.

**ideogram-oss/ideogram4** — Ideogram 4: Open image model at the forefront of design

- Repository: https://github.com/ideogram-oss/ideogram4
- Stars: 2,865 · Forks: 285
- Language: Python
- License: Apache-2.0
- Published: 2026-09-17 · Updated: 2026-09-17 · Language: en
- Canonical page: https://hysenlabs.com/projects/ideogram-oss-ideogram4

## The code is Apache-2.0 and the weights are not

The most important fact about this release is the licence split, and it is easy to miss because the repository licence is the permissive one.

The repository carries an Apache 2.0 licence file, and the package metadata points at it. The weights do not. Both entries in the model table are licensed under a document with a different name, an Ideogram 4 non-commercial licence, and that document lives in its own directory in the repository rather than at the root.

So the inference code is permissively licensed and the model is not. Someone building a commercial product can fork the Python package freely and cannot ship the weights under the same terms, which is a different proposition from an open-weight model in the usual sense of that phrase.

The project is careful about the framing elsewhere. It calls this the company's first open-weight model, and the readme opens by saying openness drives innovation and inviting the research community. It also states plainly that the model is a foundation model trained from scratch rather than a fine-tune of an existing one, which is the substantive claim behind the release rather than the licensing.

The licence question is the one to settle first, because it decides whether the rest of the setup is worth your time.

## The weights are gated, and a refused download looks like a missing file

Access to the weights is gated on the model hub, and the documentation explains exactly how that fails.

You open the model page for either build, accept the licence gate, create an access token, and then authenticate so the download carries it:

```bash
hf auth login
```

Alternatively the token can be exported as an environment variable.

The failure mode is documented as a 404 or a gated repository error. That is worth internalising, because a 404 on a repository that demonstrably exists sends people looking for a wrong path or a typo in the repository name rather than at their own account. The download is not missing; the request is unauthenticated.

Two builds are published. One is quantized to four bits and the other to eight bits, both at the same 9.3B parameter count. The project says it plans to support more quantizations.

So the access sequence is: accept a licence gate you should read, create a token, authenticate, then pull. Nothing about it is automatic, and each step is a place where the process stops with an error message that does not name the real cause.

## The default prompt path calls Ideogram's API and reads an API key

Here is the fact that changes what open-weight means for this model in practice.

The command line accepts a plain text prompt. That prompt is rewritten into the structured JSON caption the model actually expects, and the rewriting is done by what the readme calls a magic prompt model. By default, that expansion is not local. It uses Ideogram's hosted magic-prompt API, the documentation says the expansion happens server side, and it needs no local model and no local system prompt.

It reads an environment variable holding an Ideogram API key, and the key comes from the company's own developer platform.

So the default experience is: install an open-weight package, authenticate to the hub for the weights, and then send your prompt to a hosted service to be expanded. Three of the four steps are local and one is not, and the one that is not is the step that turns a sentence into the structured input the model depends on.

The repository does contain system prompt text files, which are packaged into the source distribution under the magic prompt module, so a local path exists in the shape of the code. Whether the readme offers it as a supported alternative is a different question, and the visible quick start does not.

## Only one of the two builds has Diffusers support, and it needs CUDA

The model table has a hardware column and a support column, and the two rows disagree in a way that decides your setup before you download anything.

The four-bit build lists CUDA as its supported hardware and yes for Diffusers support. The eight-bit build lists all hardware and no for Diffusers support.

So the build you want for working with the Diffusers ecosystem is also the build that will not run without an Nvidia GPU, and the build that runs anywhere does not plug into Diffusers. There is no row that is both portable and integrated.

The library behind four-bit quantization is a hard dependency of the package, which is the mechanism that makes the four-bit build work. It is listed unconditionally rather than as an extra, so even a machine taking the portable eight-bit path installs it.

The rest of the dependency list is conventional for this kind of model: a deep learning framework at a recent minimum version, the transformers library, a safe serialisation format, an accelerator library, a tensor operation helper, a sentencepiece processor, imaging, a hub client, and an HTTP client.

Python 3.10 or newer is required, and the classifiers stop at 3.12. The package installs from the repository root, and an editable install is offered for working on the code:

```bash
pip install -e .
```

## Every leaderboard chart is an image, so the positions are asserted rather than shown

The performance section is the longest part of the readme and it is built on third-party evaluations, which is the right way to make the case. But almost every number behind those positions is a picture.

The Design Arena section describes the project as the top-ranked open-weight model on the overall board, trailing only proprietary models, and as leading by a commanding margin once the board is filtered to open-weight entries. Both of those charts are images.

The same applies to the LMArena section, the internal evaluation, and the open-source benchmark section. The only figures that appear as text are the ones from the typography study, and those are specific: a blind evaluation run by ten professional designers, in which the model was picked as best of four 47.9 percent of the time, against 30.0 percent for one competitor and around 15 percent for two others, plus a usability rating of 3.55 out of 5 against 2.84, 2.61, and 2.49.

So the strongest claims in the section are the ones you can read, and they concern typography rather than overall quality. If a ranking matters to your decision, the charts have to be fetched from the linked pages rather than taken from this document.

## Text rendering is the claim the parameter count argues against

One comparison in the readme is worth pulling out, because it is the one place where a number is given against larger models.

At 9.3B parameters, the readme says the model delivers the best text rendering of any open-weight release it benchmarked, ahead of models at 20B, 32B, and an 80B mixture-of-experts model. So the argument is that a model roughly a quarter the size of the 32B competitor and a ninth the size of the 80B one renders text better.

That is consistent with the rest of the design emphasis. The model introduces a structured JSON prompting interface, and the features it names are layout and text rather than photographic realism: multilingual text rendering, explicit bounding box layout, colour palette controls, and native two thousand pixel resolution.

The internal evaluation tells the same story from the other side. Graphic designers familiar with professional design work rated blind, on graphic design and photography, and the Bradley-Terry scores put the model second overall behind one proprietary model while first among open-weight ones.

Its own open-source benchmark results name the four axes it claims to close the gap on: layout control, spatial reasoning and object fidelity, text rendering, and prompt alignment.

## Two space indentation, a committed licence directory, and no tagged releases

A few repository details tell you what kind of artefact this is.

The Python code is indented with two spaces rather than four, and the lint configuration says so explicitly, with a comment that the limit is set to match that indentation so two style checks do not fire. That is a deliberate house style rather than a mistake, and it is the kind of thing that shows up in the first file you open.

The tree is small and legible: a pre-commit configuration, the root licence, the readme, an assets directory for the logo and the chart images, a documentation directory, the separate directory holding the model licence, the project metadata, a single inference script at the root, and the source directory. There is one entry point rather than a package of commands.

There are no tagged releases. The release is announced in the readme's news section with a date of 2026-06-03, and the branch was last pushed on 2026-06-30. So the model has a launch date and a commit history rather than a version history, and anyone pinning should pin a commit or a container of their own.

The documentation site and the developer platform are both linked from the front page, alongside a collection page on the model hub.

## Conclusion

Ideogram 4 suits a researcher or a designer who wants to run a text-to-image model locally with real layout and text control, and who is willing to accept a non-commercial licence on the weights and a network dependency for prompt expansion. It does not suit a product that needs commercial rights in the model, and it does not suit an offline pipeline as shipped. Before you commit, check three things: whether your use is commercial, because the weights are not; whether you have CUDA, because the quantized build that has Diffusers support requires it; and whether you need the prompt expansion to run locally, because the default path sends your prompt to Ideogram's hosted service and reads an API key.

## FAQ

### what is ideogram 4

A 9.3B parameter open-weight text-to-image foundation model from Ideogram, released on 2026-06-03 and trained from scratch rather than fine-tuned from an existing model. It uses a structured JSON prompting interface and offers multilingual text rendering, explicit bounding box layout, colour palette controls, and native 2k resolution output. Two builds are published, one four-bit quantized requiring CUDA with Diffusers support, and one eight-bit that runs on any hardware without it.

### how to install ideogram 4

Run pip install . in the repository, or pip install -e . if you intend to modify the code so changes under the source directory take effect without reinstalling. The weights are gated on the model hub, so you must accept the licence gate on the model page, create an access token, and authenticate with hf auth login or export a token variable, otherwise the download fails with a 404 or gated repository error. Python 3.10 or newer is required.

### Can I use Ideogram 4 commercially?

Not with these weights. The repository code is Apache 2.0, but both published model builds are licensed under an Ideogram 4 non-commercial licence held in a separate licence file in the repository. If your use is commercial you would need different terms from Ideogram; the permissive licence covers the inference code, not the model.

### What does it cost to run Ideogram 4 locally?

The repository does not publish a price for local inference, and it is not an API you pay per image locally. The cost that is documented is a dependency: by default the plain prompt is expanded into the structured JSON caption by Ideogram's hosted magic-prompt API, which is described as free and runs the expansion server side, and it reads an Ideogram API key obtained from the company's developer platform.

## Sources

- [ideogram-oss/ideogram4 on GitHub](https://github.com/ideogram-oss/ideogram4)
- [Issues](https://github.com/ideogram-oss/ideogram4/issues)
- [License: Apache-2.0](https://github.com/ideogram-oss/ideogram4/blob/main/LICENSE)
- [README](https://github.com/ideogram-oss/ideogram4/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/ideogram-oss-ideogram4