Model GolfGolf

Yes, you may fine-tune or adapt existing open-source models as long as the final parameter count is under 100 million and all code is open source. You must document your starting point clearly in your submission.

Submissions must include a parameter count script that counts every trainable parameter. Organizers verify counts and reserve the right to reject submissions that creatively circumvent the limit (e.g., large embedding tables with sparse usage).

All training code, model weights, configuration, and a technical report must be publicly available under an OSI-approved license. We accept MIT, GPL, Apache 2.0, and AGPL licenses.

Round one runs June 1–30, 2026. Submissions close at 23:59 UTC on June 30. Join the Discord for early-access resources.

The model can be trained to speak clearly in one or more languages. Translation to your preferred language can be handled separately by you or your application, allowing the model to focus on its core task while you adapt the output as needed.

Models can be either pre-trained (sentence continuation) or post-trained (Q&A style). Pre-training teaches the model language patterns from vast data, while post-training fine-tunes it for specific formats like instruction following and conversational responses. Both approaches are valid for Model Golf.