SWE-bench
swebench.com
- Free plan
- Windows
- macOS
- Linux
- Web
- API
- Self-hosted
What SWE-bench is
SWE-bench is a free benchmark for evaluating language models on software issues collected from GitHub. A model generates a patch, and the evaluation harness applies it in Docker before running repository tests to check whether the issue is resolved. The original dataset contains 2,294 tasks from 12 Python repositories. Related datasets include Lite, Verified, Multimodal, and Multilingual. Verified has 500 engineer-checked solvable instances; Multilingual has 300 instances across nine languages and 42 repositories; Multimodal has 100 development instances with screenshots and interface elements, evaluated through the SWE-bench API. Its CLI submits predictions, retrieves reports, manages runs, and tracks cloud performance. Evaluations can run locally or through Modal and sb-cli. Installation requires Python 3.9 or later and Docker. Guides cover macOS, Linux, and Windows, with WSL 2 required on Windows.
Who it suits
It suits researchers and developers evaluating language models on real software issues. The datasets, CLI, local Docker setup, and cloud options support different evaluation workflows.
Good to know, and what to watch
- Uses real GitHub software issues
- Multiple benchmark variants available
- CLI manages predictions and reports
- Local and cloud evaluations supported
- Free to use
- Docker is required for consistent evaluation
- Python 3.9 or later is required
- Setup recommends 120 GB disk and 16 GB RAM
- Windows setup requires WSL 2
The verdict
SWE-bench provides structured datasets and tooling for testing model-generated patches against repository tests. Its variants and CLI support local or cloud evaluation, but Docker and substantial resources are part of the setup.
What the maker publishes
- Benchmark variants
- The site lists SWE-bench Lite, Verified, Multimodal, and Multilingual datasets alongside the original benchmark. source
- CLI
- SWE-bench CLI can submit model predictions for evaluation, retrieve reports, manage evaluation runs, and track performance in the cloud. source
- Cloud evaluation
- Evaluations can run in the cloud using Modal, or be submitted through the sb-cli tool. source
- Cloud evaluations
- The documentation describes cloud-based evaluations using Modal and the sb-cli tool. source
- Dataset variants
- The project lists SWE-bench, SWE-bench Lite, SWE-bench Verified, and SWE-bench Multimodal as available datasets. source
- Evaluation limit
- Docker is required for consistent evaluation environments, according to the FAQ. source
- Evaluation limits
- The guide notes that caching all Docker images can require about 2,000 GB of storage. source
- Evaluation method
- The evaluation harness applies a model-generated patch in a Docker environment and runs repository tests to check whether the issue is resolved. source
- Evaluation method
- Evaluation applies generated patches to repositories and runs tests in Docker containers to check whether issues are resolved. source
- Founded
- 2023 source
- Installation
- The maker's installation guide requires Python 3.9 or later and Docker, and says SWE-bench is designed to run on most systems with Docker support. source
- Installation
- The project requires Python 3.9 or later and Docker for the evaluation environment. source
- Integrations
- The project documents using Hugging Face datasets to load SWE-bench data and Modal for cloud evaluations. source
- Multilingual dataset
- SWE-bench Multilingual contains 300 instances spanning 9 languages and 42 repositories. source
- Multilingual dataset
- SWE-bench Multilingual contains 300 tasks from 42 repositories across nine programming languages. source
- Multimodal dataset
- SWE-bench Multimodal has 100 development instances with screenshots and UI elements, and test evaluation is on the SWE-bench API. source
- Multimodal dataset
- SWE-bench Multimodal contains 480 instances with issues described using visual elements. source
- Operating systems
- The Docker setup guide provides installation instructions for macOS, Linux, and Windows, with WSL 2 required on Windows. source
- Original dataset
- The original benchmark contains 2,294 task instances drawn from 12 Python repositories. source
- Original dataset
- The original benchmark contains 2,294 task instances drawn from pull requests and issues in 12 Python repositories. source
- Purpose
- SWE-bench evaluates language models on real-world software issues collected from GitHub by asking them to generate patches. source
- Purpose
- SWE-bench evaluates language models on real-world GitHub software issues by asking them to generate patches that resolve the issues. source
- Resource requirements
- The Docker guide recommends at least 120 GB of free disk space and 16 GB or more of RAM. source
- Support
- The contact page directs general inquiries to the SWE-bench team by email. source
Checked against the maker's pages on 9 Oct 2026.