We built an open-source model that cuts AI agent token bills by 74% — now we're giving away $630 to whoever builds the leanest thing with it

Every AI agent ships mountains of context to the model on every request, and most of it is noise. You pay for it, you wait for it, and it doesn't make the output any better.

So we built Paritok — an open-source (Apache 2.0) compression model trained specifically on real coding-agent trajectories. It drops in as middleware between your agent and the LLM API and cuts input tokens by ~74% on a real SWE benchmark, while keeping ~86.5% of solve quality. No rewrites: point Claude Code, Cursor, or any BASE_URL-respecting agent at it and go.

Now we want to see what people build with it, so we're running The Token-Efficiency Hackathon:

$630 in prizes

Build any LLM-powered project — agent, dev tool, RAG app — and use Paritok to make it lean