Three years ago, city names on AI-generated travel posters were often wrong—so wrong that people turned them into memes. Mohammad Norouzi remembers that at the time only a few people were building image models, and the first question facing these teams was: what makes us different from everyone else. (a16z interview[1])
In the a16z interview published on June 15, 2026, Ideogram’s founder and CEO talked about this period again. The host introduced the company, which is headquartered in Toronto, as having just released its first open-weight image model, while Ideogram’s earlier models had all been closed-source. According to the host, the new model can generate images at up to 2K, has precise layout control, and can accurately render paragraphs of text; the text can either be provided by the prompt or generated by the model itself.
What Norouzi confirmed in the interview was another set of things: prompts can run to thousands of words, and the position of each element in the image, the area it occupies, and the number of elements can all be specified. These parameters sound trivial, but when it comes to making images they are very concrete—where the text goes, how big it is, and how many there are have to be decided by the person making the request, not left to the model to improvise. (a16z interview[1])
Image: Source Ideogram actual image source[2], obtained on 2026-11-10, original interface preserved.
He put the need plainly. For use cases such as design and marketing, what is really wanted is editable design, not a flat image; the editable text and layout control he personally most looked forward to had not yet been released at the time of this interview. According to him, this open-weight release is the foundation for later features. (a16z interview[1]) A founder willing to say in a public interview that the feature he most looked forward to had not yet shipped says more about where the product stood at the time than any talk of “disrupting design”: what it delivered first was a foundation on which models could run.
Norouzi’s reasons for open weights were practical: expanding partnerships with inference providers and large enterprises. Companies can customize the models themselves, deploy them locally, or optimize them for devices; he also hopes to use open weights as an entry point for reaching partners such as chipmakers and application developers. (a16z interview[1]) He also mentioned that he hopes every artist can customize the model to fit the nuances of their own style and the texture of their canvas, so that the model becomes part of their workflow. (a16z interview[1])
The company’s own form is also shifting. Norouzi said Ideogram used to work on first-party applications, a first-party API, and model research at the same time, and now it has decided to focus more on the model side while still wanting to retain direct interaction with users. (a16z interview[1]) The center of gravity is moving from applications to models, but it does not want to cut off direct contact with users entirely—after all, feedback still has to come from users.
Another company, Recraft, has built editing into its product: a generated result can be double-clicked to enter shape editing, its anchors can be dragged, colors changed, and elements copied; when finished, it can be exported as SVG. (Recraft vector editor[3])
In the June 2026 interview, what Ideogram delivered was an open-weight image model that lets enterprises customize and deploy it in their own environments; the editable text and layout control the founder most looked forward to had not yet been released. Recraft’s vector editor page shows vector graphics that can be edited directly; what the two sides deliver is not the same kind of thing. Open weights can become an entry point for partnerships with inference providers and large enterprises; after enterprises get the model, they can customize it themselves and deploy it locally. Whether the editable design he hopes for can land in workflows as he envisions is not entirely up to him.