Fusions are more capable than any other AI system in the world, both in absolute terms and capability/price. Now you can see it.
Research already shows multi-model compositions (routed ensembles, cascades, councils, multi-agent systems) show increased capabilities at the same cost.
But each result lives in its own code, on its own benchmark, against last year's models, and testing against today's frontier means paying the full inference bill before seeing a single number.
The ScreamingFace leaderboards centralize this prior work, enabling researchers to test it against today's frontier, move the frontier forward, while allowing everyone to see and reproduce them at no cost.
Every fusion entry comes with its code, in a single url4 line that contains all the details of the run. Each call in the recipe is pinned to an exact model version, prompt, and seed.
A result appears only after it can be independently rerun on the hosted engine and registered on the shared cache. Replaying a published row costs nothing, and starting from someone else's recipe means you pay only for what you change.
Each leaderboard plots every fusion and every solo model by score and run cost, and the frontier is the edge of that plot, revealing the best accuracy anyone has reached at each budget. Fusions and the strongest solo models sit on the same board, so you can see exactly how each fusion compares.
A fusion depends on its member models, but not on any single one. If a developer reprices or retires a model, the board shows which replacements come closest and what each one costs.
Each leaderboard covers one benchmark. Put together, they form the first verified map of the quality–cost frontier, showing where combinations of models beat the best single model and where they don't yet. Verified means a result only appears on the leaderboard after it has been run on the shared engine with full cached results and confirmed reproducibility, so anyone can re-run and obtain the same results.
Anyone can compose a fusion using the Client, compile it to a url4 recipe, score it against a benchmark of your choice, and publish it on its corresponding leaderboard. Code and recipes are open under Apache-2.0.
ScreamingFace is built in the open by a community of researchers and engineers, with the support of OpenMined.