Score is training an open-source world model on Bittensor Subnet 44, releasing it at every stage and selling access through its own platform. The team chose that target over a vision-language model, saying the path to a useful VLM is mapped well enough now that another one adds little to open-source progress.
Subnet 44 has already produced evidence the mechanism works. A 19MB detection model that came out of the subnet's competition scored 0.848 mAP on the UA-DETRAC vehicle benchmark in June, running on a four-thread CPU while beating OWLv2, GPT-4o, Gemini, Grok, Claude, SAM3, Grounding-DINO, and DETR.
Score arrived at this point through a failed first product, a rebuilt incentive mechanism, and a commercial platform still on a waitlist. The world model is the next step and the only one not yet built.
