Beat the best single model with a fusion.
Prove it on a public leaderboard.
No single model wins every benchmark. Fusion Monsters is a program for researchers who want to test that seam: build a fusion of models, post a reproducible score, and put your name on the result. The application takes about five minutes.
Want to read ahead? The challenge, in brief ↓
the monsters, already at work — every reproducible win joins the cast
What Fusion Monsters is
A supported technical challenge for people building and testing model fusions — combinations of models that work together on a benchmark. Applicants selected after review receive a challenge brief with the benchmark, the permitted models, access details, and a deadline. The cohort is then selected from the application, the challenge attempt, and the direction you'd take with a larger budget.
- Apply. A short profile, links to your technical work, availability, and what you want to explore. No essays, and nothing to sign — the full rules of the game arrive if you join, not before.
- Review. The program team reads for relevant experience, fit, and capacity.
- Challenge. Selected applicants get $50 of credits and a fixed benchmark task: build a fusion from open models and beat the baseline we set with our own strong attempt. Short terms come with the credits; you accept them before the credits are used.
- Decision. Admissions are rolling and nobody is flat-rejected — the strongest attempts enter first, everyone else keeps their place in line.
The challenge is not pass-or-fail on one score. A serious attempt, clear reasoning, and a useful next-step plan all count.
Who it's for
Researchers and engineers who already run the loop: find a benchmark, study the current best, try something smarter, publish the number. Fusion Monsters gives that loop a permanent home — a leaderboard that stays live, recipes anyone can fork, and compute so cost is never the reason you stop.
If you'd rather explore quietly than publish, the tools don't require the program: everything runs on your own keys, and the leaderboard is public either way.
What winning means
Every reproducible entry lands on the accuracy-versus-cost chart next to the strongest single models. A win is a spot on that frontier: a reproducible score above the best published result, or the same accuracy at a cost nobody else has reached. Reproducible means the score came from a full run of your recipe on the stack, and re-runs to the same number — a score nobody has to take on faith.
Winning means publishing. Your recipe goes up openly, under your name. The next Monster starts from it and tries to beat you. That's the point.
Apply
Applications open at program launch. Leave your email and the application form lands in your inbox the moment it does. It takes about five minutes, and there's nothing to sign.
Done. The application comes to you at launch.
That didn't go through. Check the address and try again.
Questions people ask before applying
What counts as a win?
A reproducible spot on the accuracy-versus-cost frontier for a supported benchmark: a score above the best published result, or matching accuracy at the lowest full-run cost. Reference baselines — the strongest single models — sit on the same chart.
What does "reproducible" mean?
The score came from a full run of the recipe through the program's stack, and it re-runs to the same number. Results first run elsewhere can be submitted, but they appear labeled self-reported until an on-stack re-run succeeds. A reproducible entry always ships with the configuration needed to reproduce it.
What do members get?
A monthly compute allowance ($300 to start) for building and sample-running fusions, access to the supported benchmarks and models, and the program's support channel. A strong reproducible result can earn a celebration grant — a raised allowance for the months that follow.
Can I use my own API keys?
Yes — your own keys and models beyond the supported set, at your own cost. For a result to publish as reproducible, the run still goes through the platform.
When do I agree to the rules?
Not at application — applying commits you to nothing. Short terms arrive with the challenge credits and you accept them before the credits are used. The full rules of the game come at the join step, before your member allowance is issued. Published recipes carry the Apache 2.0 license under the leaderboard's terms.
Do I have to publish my recipe if I win?
Yes. A reproducible win is published with its recipe under your name, under the Apache 2.0 license. If you want to keep a method private, don't submit it for an official win.
What are the rules on fair play?
Real, reproducible recipes; opaque components declared; no probing benchmarks for leaks. A result can be held for closer review — that's not an accusation, it's what keeps the board worth winning.