Caffin/SVGThinker-7B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 1, 2025License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Caffin/SVGThinker-7B is a 7.6 billion parameter text-to-SVG generation model fine-tuned from DeepSeek-R1-Distill-Qwen-7B, designed to produce editable SVG code from natural language descriptions. It specializes in generating compact, icon-style vector graphics, focusing on structured code generation rather than raster images. The model has a context length of 32768 tokens and is optimized for creating SVG icons and simple vector graphics from English text prompts. It achieves a FID score of 34.06 and a CLIP score of 0.2765 on its paper's held-out text-to-SVG prompts.

Loading preview...

SVGThinker-7B: Text-to-SVG Generation Model

SVGThinker-7B is a 7.6 billion parameter model developed by Caffin, specifically engineered for text-to-SVG generation. Fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-7B, this model translates natural language descriptions into editable SVG code, with a particular emphasis on creating compact, icon-style vector graphics. Its training directly in SVG code space, informed by a sequential annotation pipeline, helps it generate more editable and structured SVG.

Key Capabilities

  • Generates editable SVG code from English text prompts.
  • Specializes in icon-style vector graphics and simple prototypes.
  • Focuses on structured code output rather than raster images.
  • Supports a 32768-token context length for detailed prompts.
  • Achieves competitive performance with a FID score of 34.06 and a CLIP score of 0.2765 on its dedicated text-to-SVG evaluation set.

When to Use This Model

SVGThinker-7B is ideal for:

  • Rapid prototyping of simple vector graphics and icons.
  • Generating SVG assets for web development or UI/UX design.
  • Research into text-to-code generation, particularly for structured graphic formats.

Limitations

Users should be aware that outputs may occasionally be malformed or visually inconsistent. The model is best suited for simple to moderately complex icon-style graphics and may struggle with photorealistic scenes, dense layouts, or text-heavy SVGs. Generated SVG should always be reviewed and sanitized before production use due to its executable nature.

Finetunes
1 models