Benchmark
SVGBench
Coding benchmarks compare models on programming tasks. Scores are percentages. Higher is better.
- Models measured
- 82
- Scores recorded
- 88
- Sources
- 3
- IQ weight
- 2.21
Score distribution
- 47.1%
- Mean
- 48.2%
- Median
- 14.6%
- Std dev
- 28.4%
- 10th percentile
- 64.1%
- 90th percentile
- 6.8%
- Min
This benchmark has a weight of 2.21 for chat models, measured over 82 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | GPT 5.2svg-generation, xhigh· SVGBench data | 74.4% |
| 2 | Claude Opus 4.5svg-generation, non-thinking· SVGBench data | 72% |
| 3 | Claude Opus 4.5svg-generation, thinking· SVGBench data | 71.5% |
| 4 | GLM 5svg-generation· SVGBench data | 70.3% |
| 5 | Gemini 3 Pro Previewsvg-generation· SVGBench data | 68.7% |
| 6 | GPT 5.1svg-generation, high· SVGBench data | 67.5% |
| 7 | Kimi K2.5svg-generation· SVGBench data | 67.5% |
| 8 | GPT 5svg-generation, high· SVGBench data | 67.4% |
| 9 | MiniMax M2.5svg-generation· SVGBench data | 65% |
| 10 | Gemini 3 Flashsvg-generation· SVGBench data | 64.1% |
| 11 | MiniMax M3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 63.7% |
| 12 | MiniMax M3official model-page overview figure· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 63.7% |
| 13 | MiniMax M3text/image inputs · VLM rendering verifier · three-run mean· Reported by MiniMax Research· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 63.7% |
| 14 | MiniMax M2.1svg-generation· SVGBench data | 63.1% |
| 15 | GLM 4.7svg-generation· SVGBench data | 62.4% |
| 16 | Claude Sonnet 4.5svg-generation· SVGBench data | 62.2% |
| 17 | Gemini 2.5 Prosvg-generation· SVGBench data | 61.4% |
| 18 | GPT-5 Codexsvg-generation, medium· SVGBench data | 61% |
| 19 | GLM 4.6svg-generation· SVGBench data | 60.5% |
| 20 | Claude Sonnet 3.7svg-generation· SVGBench data | 60.4% |
| 21 | GPT-5 minisvg-generation· SVGBench data | 59.7% |
| 22 | Claude Sonnet 4svg-generation· SVGBench data | 59.4% |
| 23 | GPT 4.1svg-generation· SVGBench data | 58.4% |
| 24 | GLM 4.5svg-generation· SVGBench data | 58.3% |
| 25 | MiniMax M2svg-generation· SVGBench data | 58.3% |
| 26 | Claude Haiku 4.5svg-generation· SVGBench data | 57.2% |
| 27 | Kimi K2 Thinkingsvg-generation· SVGBench data | 57.1% |
| 28 | OpenAI o3svg-generation· SVGBench data | 56.7% |
| 29 | Grok 4svg-generation· SVGBench data | 54.4% |
| 30 | GPT-4.1 minisvg-generation· SVGBench data | 53.4% |
| 31 | DeepSeek V3.1svg-generation· SVGBench data | 53.1% |
| 32 | gemini-2.5-flash-preview-09-2025svg-generation· SVGBench data | 52.1% |
| 33 | Gemini 2.5 Flashsvg-generation· SVGBench data | 51.4% |
| 34 | DeepSeek V3.2 Expsvg-generation· SVGBench data | 51.3% |
| 35 | Grok 4 Fastsvg-generation· SVGBench data | 50.5% |
| 36 | GPT-5 Chatsvg-generation· SVGBench data | 50.4% |
| 37 | DeepSeek V3.2svg-generation· SVGBench data | 49.8% |
| 38 | Qwen3 Coder Plussvg-generation· SVGBench data | 49.8% |
| 39 | GPT-5 nanosvg-generation· SVGBench data | 49.7% |
| 40 | GLM 4.5 Airsvg-generation· SVGBench data | 48.9% |
| 41 | Qwen3 Codersvg-generation· SVGBench data | 48.7% |
| 42 | Qwen3 235B A22B Instruct 2507svg-generation· SVGBench data | 48.6% |
| 43 | Kimi K2 0905svg-generation· SVGBench data | 48.5% |
| 44 | DeepSeek V3.1 Terminussvg-generation, reasoning· SVGBench data | 48.2% |
| 45 | o4 Minisvg-generation· SVGBench data | 48% |
| 46 | Claude 3.5 Sonnetsvg-generation· SVGBench data | 47.9% |
| 47 | DeepSeek V3.1svg-generation, reasoning· SVGBench data | 47.8% |
| 48 | Qwen3 Next 80B A3B Instructsvg-generation· SVGBench data | 47.4% |
| 49 | GLM 4.5Vsvg-generation· SVGBench data | 46.5% |
| 50 | GPT-5.1 Chatsvg-generation· SVGBench data | 46.3% |
Showing 1–50 of 88 results.
Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats