Independent research on local AI and open source.
Controra Research is an independent initiative working on local AI and open source. We aim to publish our research on GitHub and to run benchmarks with the resources available to us.
Human knowledge was built collectively over thousands of years, and that shared inheritance should be accessible. Intelligence should not be exclusively owned. AI is science, and science should be open. People should be able to contribute, and many independent contributors can make a difference. We pursue local AI and open source so more people can take part.
We work on local AI research, experiments and tooling. Alongside this, we plan to work on AI products meant to help people explore, learn and dream bigger. Some of these products may be closed source and could help fund the research while we look for other ways to sustain the initiative.
Sparse Decode Bench investigates when sparse decoding over pages of the key/value cache pays off for ordinary dense language models in llama.cpp on consumer NVIDIA GPUs, weighing page selection overhead, decoding speed and model quality.
Public repositories on GitHub.
Pins Hugging Face model revisions in a lockfile and proposes updates as GitHub pull requests. An optional validation script can test each candidate before its pull request opens.
A local dashboard that shows coding agent sessions side by side, with each session's status and a button that takes you to it.