SWE-bench Multimodal (Vals v1): Carbon
Code
Vals v1 issue-resolution rate on its Carbon repository panel; one offline mini-swe-agent run with screenshot attachments and official test grading. Subset size is not published.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 52.6 |
| Claude Fable 5.1 | 44.4 |
| Gemini 4 Argon | 43.6 |
| Claude Sonnet 5.5 | 42.9 |
| Kimi K3 | 30.1 |
| DeepSeek V4.1 Flash | 29.3 |
| Muse Wonder Full | 29.3 |
| Muse Wonder | 27.1 |
| MiMo-V2.6-Pro | 26.3 |
| Step 5 Preview | 24.8 |
| Claude Haiku 5.5 | 24.1 |
| GPT-6.1 Sol | 24.1 |
| GPT-6 Astra | 24.1 |
| Grok 4.7 | 24.1 |
| Muse Spark 1.3 | 22.6 |
| Mistral Large 4 Preview | 20.3 |
| Gemini 3.8 Flash | 19.5 |
| GPT-6 Luna | 19.5 |
| MiniMax M3 | 19.5 |