Arrow 2 transforms every generation into an editable SVG

QuiverAI is expanding its family of vector models with Arrow 2 and Arrow 2 Telos, capable of generating, vectorizing, modifying, and animating SVGs from an application or an API.

A PNG image loses sharpness when enlarged. An SVG can be scaled without that problem because the image is described through shapes, paths, colors, and coordinates. QuiverAI is applying generative AI directly to this editable format with Arrow 2, a new family of models designed to create, convert, modify, and animate vector graphics.

Arrow 2 is presented as the company’s largest update since the launch of its first vector-generation model. It accepts written instructions and visual references, then produces an SVG file that can be opened, inspected, and modified in conventional design or development tools.

This distinction separates Arrow from image generators that merely imitate vector artwork. A conventional generator may produce a flat illustration with clean edges, but its output is still a raster image. Arrow generates the underlying paths and groups that make up the graphic.

The result can therefore be resized, recolored, rearranged, or integrated into a website without first being traced manually. A designer can adjust individual shapes, while a developer can manipulate the same file through CSS, JavaScript, or another interface that supports SVG.

QuiverAI says Arrow 2 produces cleaner geometry than the previous version. Its files are intended to use fewer and more precise control points, reducing unnecessary complexity, intersecting paths, and irregular contours.

A lower number of control points does not automatically guarantee a better drawing. It can, however, make a file easier to edit and reduce the risk that a simple curve is represented by dozens of small segments. This matters when the result must be refined by a person rather than merely displayed once.

The model also aims to improve composition. QuiverAI says it pays more attention to spacing, padding, alignment, and the relationships between the different parts of an image. These properties are particularly important for diagrams, interface assets, icons, and illustrations that must remain readable at several sizes.

Arrow 2 can generate an image from a written brief or use an existing visual as a reference. In the latter case, the model attempts to preserve the palette, shape language, and graphic treatment while producing new variations.

A team could use this capability to create several characters in the same visual family, expand a set of icons, or explore alternative versions of an existing brand asset. The reference guides the result, but it does not provide a formal guarantee of perfect consistency across an entire collection.

The examples published by QuiverAI include character portraits, anatomical diagrams, fashion drawings, and technical line work. These cases require more than a visually attractive result. Seams, closures, highlighted regions, labels, and repeated structural details must remain coherent from one variation to the next.

The company does not publish a quantitative evaluation of these improvements. There is no benchmark measuring the number of malformed paths, the accuracy of diagrams, consistency across several generations, or the amount of manual correction required before a file can be used professionally.

Claims about higher quality and faster output therefore come primarily from QuiverAI’s own examples. The demonstrations show more controlled compositions and cleaner files, but they do not establish how consistently the model performs across different styles, instructions, or levels of complexity.

Arrow 2 also handles vectorization. A user can provide a sketch or raster image and ask the model to recreate it as an editable SVG. Instead of simply placing the original image inside an SVG container, the service attempts to reconstruct its visible elements as vector shapes.

This process can be useful for converting hand-drawn concepts, old logos, product illustrations, or flat graphics into reusable assets. The result still needs to be reviewed. Fine textures, shadows, lettering, gradients, and ambiguous contours may be simplified or interpreted differently from the source.

Vectorization also raises a structural question that cannot be assessed from appearance alone. Two SVG files may look identical while being organized very differently. One may contain clearly named and reusable groups, while the other may consist of a long series of paths that are difficult to understand or edit.

Arrow 2 extends beyond static graphics through micro-animations. It can animate the shapes and groups already present in an SVG to create logo reveals, loading indicators, animated icons, product illustrations, and other interface elements.

Working from the existing structure can help motion remain connected to the design. An eye can blink, steam can rise from a cup, or parts of a symbol can appear in sequence without the asset being converted into a conventional video.

This does not make Arrow 2 a complete animation environment. Complex timelines, physical simulations, interactions, and production controls may still require dedicated software or additional code. The feature is primarily suited to short, structured motion built from the components of an SVG.

QuiverAI is releasing a second model called Arrow 2 Telos for more demanding work. Telos combines Arrow’s vector infrastructure with the refinement capabilities of larger frontier models. It is intended for briefs involving more elaborate styles, compositions, and requirements.

The company recommends Arrow 2 when speed and cost per asset are the main priorities. Telos is positioned for tasks that may benefit from additional interpretation and refinement. This creates a choice between a faster model for routine production and a more resource-intensive option for complex creative work.

According to the model documentation, Arrow 2 uses the identifier `arrow-2` and supports a context window of 131,072 tokens. Arrow 2 Telos uses `arrow-2-telos` and expands that capacity to 1,050,000 tokens. Both can generate up to 65,536 output tokens.

The large context window can accommodate detailed