Quiver AI raises $8.3M to generate SVGs through code — not tracing — using PhD-trained models
Mar 2, 2026 with Joan Rodriguez
Key Points
- Quiver AI raises $8.3 million to generate SVGs as compilable code rather than traced pixels, producing outputs directly editable in design software without manual cleanup.
- Co-founder Joan Rodriguez trained custom reinforcement learning models during his PhD to prioritize SVG code generation, betting specialization outpaces general-purpose models as the category scales.
- Quiver launches consumer website while deploying API for agents and MCP servers, signaling early traction in developer and design-tooling layers ahead of direct designer adoption.
Summary
Joan Rodriguez, founder and CEO of Quiver AI, has raised $8.3 million to build AI models that generate SVGs through code rather than the traditional tracing approach that has long frustrated designers.
An SVG is structured code that compiles into an image. Rodriguez argues that treating generation as a code problem instead of a pixel-tracing problem produces cleaner, more editable output. LLMs are already strong at code generation. During his PhD, Rodriguez trained models specifically on this task and developed custom RL reward signals to push quality further.
For designers, SVG outputs drop directly into Figma or Illustrator for final editing, with colors, shapes, and splines all adjustable. A near-perfect SVG at 99% quality is far more useful than a near-perfect PNG, which requires manual cleanup that can take significantly longer.
Launch and distribution
Quiver launched with a consumer website to observe how people use the tool while simultaneously building an API so other software can call its models directly. Agents are already being deployed against the API, and MCP servers are calling Quiver, suggesting early traction in the developer and design-tooling layer rather than purely direct-to-designer.
Competition
The big labs are moving into SVG generation. Throwing a large model at a large dataset produces good results in this space. Quiver's bet is that deep specialization through custom RL rewards and purpose-built training gets further than a general-purpose model with broad coverage. Whether that moat holds as frontier labs invest more directly in the category remains unclear.
Rodriguez is relocating the company to San Francisco from Barcelona.