Superior Ingredients presents: is a competitive league for open-model-building. SI 1, Model Chess
An open league for open intelligence.
Four divisions. Featherweight (<8b params). Lightweight (8-20b). Middleweight (20-70b). Heavyweight (>70b).
Keep intelligence open. Keep it fighting.
America is not short of open models so much as the institutions with an incentive to build them, and the American open-source coalition is mobilized yet unled. Token volume is moving toward specialized, constrained models, and post-training them has become a first-class discipline with a real market and no way to measure superiority. Meanwhile, the builders doing the actual work have never had a stage.
In the SI League, builders compete in four weight classes to build the best open models under strict constraints. The argument every competition makes is that what a model is made from, and who made it, matters more than how large it is.
Divisional fights run quarterly. The divisions are as follows: Featherweight, <8B parameters; Lightweight, 8-20B; Middleweight, 20-70B; and Heavyweight, >70B. Champions hold belts and defend them. Alongside the divisions, the league runs one-off competitions with unique formats and constraints, some designed with and commissioned by the companies who care about the task at hand.
Competitors enter alone or as teams. Fights are announced in advance and judged on the night. Winning models are released under the SI label, credited to the builder.
Scoring is run by an independent evaluation partner. Divisional fights are scored on coding tasks written fresh for each occurrence, sealed until submissions lock, and published in full afterward. Sponsors buy affiliation with the league, its competitions, and its teams. They never buy anything to do with the outcome. There are no exceptions.
The league exists to find superior ingredients, the label exists to ship them, and the competitions exist to witness the finding. This is where the good stuff is proven.
Keep intelligence open. Keep it fighting. 🇺🇸
Format. Open entry chess tournament. All submitted models play a qualifying gauntlet against each other. Top 16 by score are seeded into a single-elimination bracket: Round of 16, quarterfinals, semifinals, final. Four rounds, one champion.
Model constraints. Under 50MB of weights, any precision. Trained from scratch. No pretrained weights, no fine-tuning of an existing model. Any architecture and any training method, all disclosed.
Training data. Every entrant trains on the same provided dataset: Lichess games where both players were rated between 1800 and 2200, standard time controls, rated games only. One versioned file with a published checksum, identical for everyone. No outside game data and no distilling from an existing chess engine. Self-play reinforcement learning on top of the provided data is permitted and encouraged. Submissions must include training code, data-loading code, and run logs alongside the weights. Misdeclaring what you trained on is the one offense that results in disqualification.
Compute constraint. A fixed budget of 24 H100-hours per entrant, sponsor-provided, usable as a single card or a node. Training on your own hardware is permitted within the same time budget and must be declared at submission.
Match format. Every match is played from the standard starting position with colors split evenly, so neither model gets an advantage from moving first. Round of 16, quarterfinals, and semifinals are two games, each model playing White once. The final is four games, each model playing White twice. A win scores one point and a draw scores half. The model with more points advances. If a match is level after its scheduled games, the models play sudden-death games with colors alternating until one is decisive. Every move must be legal and delivered within one second. An illegal move, a failure to move in time, or a crash forfeits that game.
Play conditions. Every model runs on identical hardware, air-gapped, with identical inference settings. No engine assistance, no network access, no opening books, no endgame tablebases.
Outcome. No judges, no scorecards, no benchmark. You win by beating the model in front of you. The champion takes the trophy and the purse, and their weights and full method publish under the SI label that night.
Purse. $7,500 to the winner, $2,500 to the runner-up. Funded by Superior Ingredients and guaranteed. The purse passes through in full; the league takes no cut of prize money. Entry is free.
To compete. These rules are a draft and the builders who compete should shape them. To declare intent to compete, ask a question, or argue with any of the above, DM us. Competition date TBD.