Comparing SVG Generation for the top open models
Posted by omarous@reddit | LocalLLaMA | View on Reddit | 2 comments
Some of the larger models (like Llama) weren't available on OpenRouter, so I had to work with what was there.
- Best small model: Gemma 4 26B For its size, I think it had the best output. You can see it even picked blue eyes for the husky.
- Definitely useless: Llama 4 Maverick, gpt-oss-120b gpt-oss gets a point for at least painting something that resembles a dog.
- Mid-tier: MiniMax M2.7, Qwen3.6 Max, Kimi K2.6 Lots of detail, but the dog isn't well-positioned.
- Top-tier: GLM 5.1, DeepSeek V4 Pro Pretty darn close to usable.
i_am__not_a_robot@reddit
If the goal is SVG generation, I think hybrid approaches like https://ajayj.com/vectorfusion (not in this particular implementation, but in general) are the way to go.
fantomechess@reddit
You might be interested in https://github.com/johnbean393/SVGBench but I just realized it hasn't been updated in a couple months.