Recraft: Recraft V4 Vector

Recraft: Recraft V4 Vector is an image generation model from Recraft, released on 13 May 2026. It sits in the image category and is the vector variant of Recraft V4, which means its output is SVG rather than raster pixels. If you are building anything that needs scalable artwork, icons, logos or illustrations that stay sharp at any size, this is the model in Recraft's lineup aimed at that job.

The model accepts text and image inputs and produces image output. Its context window is 65,536 tokens, with a maximum output of 58,982. Generation is controlled through image parameters rather than the usual sampling controls: aspect_ratio accepts 1:1, 4:3, 3:4, 16:9, 9:16 or auto; output_format accepts svg; n sets how many images to generate, from 1 to 6; and input_references takes between 0 and 1 reference images. There are no other supported parameters listed, so the request surface is deliberately narrow.

Those capabilities point to a fairly specific set of use cases. Because the output is SVG, the results are editable and resolution independent, which suits design systems, icon sets, brand marks and web or app assets that need to scale across screen densities. The aspect ratio options cover square, landscape and portrait layouts, so you can target social formats, banners, hero images or vertical placements without post-processing crops. Passing a single reference image lets you steer style or subject while still getting vector output, which is useful when you want a consistent look across a batch of assets. Setting n above 1 gives you several candidates in one call, handy when exploring directions before committing.

You call it through api14's OpenAI-compatible API using the model id "recraft/recraft-v4-vector". Point your existing OpenAI client at the api14 base URL and send a request in the usual chat or image style, with the model id set to that string. The parameters you can rely on are the image ones: aspect_ratio, output_format, n and input_references. Since the supported parameter list is empty beyond these, avoid passing sampling options such as temperature or top_p and expect them to have no effect.

One thing to weigh before adopting it: the only output format is SVG, so if your pipeline expects PNG or JPEG, you will need a rasterisation step downstream. The reference image limit of one also means you cannot blend multiple references in a single call. If you need raster output directly, or more flexible parameter control, a different image model is likely a better fit.