kat dev
Open-source coding LLM for software engineeringAn AI coding model you can run yourself
Highlights
What you can do with it
- Generate new features from high-level descriptions with multi-step planning
- Find and fix errors, then refactor code for clarity and performance
- Write unit tests and generate documentation for an existing codebase
How to use it
- 1 Download the model from Hugging Face, choosing the 32B or 72B Exp version
- 2 Integrate it with the Transformers library or vLLM for local deployment
- 3 Use StreamLake for enterprise-level API access if you prefer a hosted option
- 4 Start coding: generate, fix, optimize, and document with the model
- 5 Deploy to production and connect it to your CI/CD pipeline
Who it is for
Pros
- Apache 2.0 licence means free use and community research
- Self-hosting keeps your code on your own infrastructure
- Benchmark results suggest strong software engineering ability
- Local deployment through common tools like Transformers and vLLM
Limitations
- Large models need serious GPU resources to run locally
- Generated code still requires review and testing
- Benchmark scores do not guarantee results on your codebase
- Open-source licence terms and model updates can change over time
What Kat Dev is
Kat Dev is a family of large language models developed by the Kwaipilot team at Kuaishou, focused on software engineering and coding tasks. It is designed to help developers with feature implementation, error fixing, refactoring, performance optimization, test generation, and documentation, and it can coordinate multiple agents for complex engineering work. The models support multi-language code and are open source under the Apache 2.0 license, so developers can download and use them for free. The 72B Exp version is reported to have scored 74.6 on the SWE Bench benchmark, which is the headline number behind its reputation in the coding community.
What you can do with it
- Generate new code features from high-level descriptions through multi-step planning.
- Detect and correct errors, then rewrite code for clarity and performance.
- Propose optimization suggestions covering algorithms and efficiency.
- Write unit tests, explain complex code, and generate documentation.
Who it is for
- Software development teams that want AI assistance inside their pipeline.
- Researchers who need an open coding model for experiments.
- Startups that want capable code generation without licence fees.
- Open-source contributors who prefer models they can inspect and modify.
What to watch out for
These are large models, and running the 72B version locally demands serious GPU memory, so check your hardware and budget before planning a self-hosted setup, because the hosted API route exists for a reason. Generated code is a draft, not a finished product, and it still needs review and testing before it ships, especially for security-sensitive paths. Benchmark scores reflect specific test sets and do not guarantee the same results on your codebase, so run your own evaluation on real tasks. Open-source licence terms and model releases can change, so follow the official repository for current conditions.
FAQ
Is Kat Dev free?
Yes. The models are open source under the Apache 2.0 license and can be used for free.
Where do I get it?
You can download the 32B or 72B Exp versions from Hugging Face.
How was it benchmarked?
The 72B Exp version is reported to have achieved 74.6 points on the SWE Bench benchmark.
Can I use it without self-hosting?
Yes. Enterprise-level API access is available through StreamLake.
Why it is free
What the official page says: for free
Free limits, trial length and features are decided by the official site and can change at any time.
Product preview
Screenshot archived from the listing page, for preview only.