Ideogram has published the code and weights for Ideogram 4, its first open-weight text-to-image model. The company says the system was trained from scratch rather than adapted from an existing model, and it is now available through a public GitHub repository and Hugging Face collections.
The release marks a notable step for the company, which describes Ideogram 4 as an image model focused on design-quality generation. According to the repository documentation, the model supports a structured JSON prompting interface, explicit controls for bounding boxes and color palettes, multilingual text rendering, deeper language understanding, and native 2,000-pixel image output.
Ideogram says the model is intended to be used both through its own website and by developers who want to experiment with the open-weight release. The company points users to its online product as the easiest way to try the model, while also providing installation instructions and inference code in the repository.
The model zoo in the repository lists two versions of Ideogram 4, including an nf4 quantized model that is tied to CUDA hardware and an fp8 version intended for broader hardware support. Both are labeled as non-commercial in the posted license terms. The company says it plans to support additional quantizations later.
The release notes indicate that the weights and inference code became public on June 3, 2026, alongside a technical blog post. A later commit on June 4 fixed an API link in the repository documentation.
Ideogram positions the launch as part of a broader push toward openness in generative imaging. In the README, the company says it believes open access helps drive innovation and invites researchers to build on the model.
The repository also includes benchmark claims comparing Ideogram 4 with other image-generation systems. Ideogram says the model performs strongly on third-party design and text-to-image leaderboards, as well as on internal evaluations involving professional designers. The benchmarks highlighted in the repository suggest the model is especially competitive in typography and layout tasks, areas that often matter in poster design, branding, and other commercial art use cases.
The company also says Ideogram 4 ranks highly on standard open-source tests for layout control, spatial reasoning, object fidelity, prompt alignment, and text rendering. In particular, the documentation emphasizes the model’s text rendering quality relative to other open-weight releases, including larger models.
Access to the weights is gated on Hugging Face. Users must accept the license terms and authenticate before downloading the files. The documentation provides setup instructions for installing the package, logging in with a Hugging Face token, and running inference from the command line.
For users who enter a plain text prompt, the system can convert it into the structured caption format the model expects through a hosted "magic prompt" service. Ideogram says that API is free to use for prompt expansion.
The public launch places Ideogram in the growing group of companies releasing open-weight image models to developers and researchers. With Ideogram 4, the company is emphasizing design-oriented generation, text handling, and layout control as the main differentiators for the model.