Listed and community engines
Anyone can add an engine to Kalsa, whether it’s a new project, a fork or an experiment. Engines that do something better than what came before, with proof anyone can check, get the Listed mark.
We read every source, but we can’t re-run every benchmark or tell real stars from bought ones. Kalsa relies on authors being honest and on users speaking up. If a number looks wrong, write to help@kalsa.io and we’ll take another look.
Community engines
Community listing opens soon. Authors will add their own engines. We won’t check them, so read their claims with care.
Listed engines
We checked that they meet every rule below. Once community listing opens, a community engine that meets them too can become Listed through a public pull request.
What every Listed engine needs
- It does something better, with public proof.
- It has at least 5 stars on GitHub.
- It uses an open-source licence approved by the OSI.
- It’s alive: at least one commit in the last six months, and not archived.
- If it’s a fork, it says which project it comes from.
- You can install it from the project’s own releases or package registry.
- It doesn’t need sudo or changes to your system. If one feature does, we show it as a trade-off.
What counts as better
Better doesn’t only mean faster. An engine can run bigger models, use less memory, support new hardware or handle longer conversations. Whatever it improves, it needs a public source and the machine it was measured on.
| Better at | The proof we need |
|---|---|
| Writing replies faster | tok/s before and after, on the same machine |
| Reading prompts faster | tok/s before and after, on the same machine |
| Starting up faster | seconds before and after |
| Using less memory | peak memory before and after, same model |
| Running bigger models | model size, the machine’s memory, and the speed it reached |
| Handling longer conversations | context length before and after, same memory |
| Serving many requests at once | total tok/s and how many requests ran in parallel |
| Running across several machines | which machines, how they’re connected, and a speed if published |
| Supporting new hardware | a measurement on it, and proof the original project can’t run there |
| Supporting a new model | a measurement, and proof the original project can’t load it |
| Giving better answers | perplexity or a public benchmark, same model |
| Using less energy | tokens per joule, measured |
How we judge the proof
- A gain under 5% is noise, so we don’t count it.
- Every gain must show what it costs. If the cost is hidden, the gain doesn’t count.
- An engine must compare itself with the project it comes from, with the same build with the feature turned off, or with a well-known engine. Beating its own older version doesn’t count.
- A different weight format counts only if the quality loss is measured too.
- Features without a number, like tool calling or a chat interface, don’t count on their own.
Where the numbers come from
Every number comes from the engine’s author, and we link the source. We don’t re-run the benchmarks ourselves. Numbers measured on a different chip appear in grey.
What “Reviewed” means
We scanned the code a fork changed from its original project, at one specific commit, and a person checked anything the scan flagged. The full report is public. Every new commit needs a new review. A review lowers the risk, but it can’t remove it. Any engine runs with your user’s permissions, and macOS doesn’t sandbox it, so install only engines you trust.
Corrections
We write the short summary on each page. If something is wrong, or you want a page removed, write to help@kalsa.io.