agents benchmark
SWE-Marathon
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
Updated Aug 7, 2026
Published models3
Registry coverage3
MetricScore
EvidenceB
SWE-Marathon leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | MA | 42.0% | 100.0% | 3 | C | |
| 2 | XA | 29.0% | 50.0% | 3 | C | |
| 3 | ZA | 13.0% | 0.0% | 3 | C |
Score distribution
Top published rows on the benchmark's original scale.
SWE-Marathon highlights
The top published results on this benchmark's own scale.
What is SWE-Marathon?
Definition and scoring fields from the benchmark registry.
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- SWE-Marathon
- Modality
- text
- Primary category
- agents
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- swe-marathon|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about SWE-Marathon.
Which model scores highest on SWE-Marathon?
Kimi K3 is currently ranked first with 42.0%.
What does SWE-Marathon measure?
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
3 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.