The first test uses Qwen3.5 0.8B, a small open model. Five verified edits have so far lifted its score from 18.6% to 19%. The gain is modest, but it shows the verification process working on a live model.
Qwen3.5 is a family of open-weight models from Alibaba. Wikipedia’s entry on Qwen notes that Alibaba released the open-weights Qwen3.5 in February 2026.
What Comes Next
Support for changing model designs and custom tests for businesses will follow in the coming days. Allowing changes to a model’s architecture, and not only its weights, would widen the range of contributions the contest can accept. Custom tests would let companies measure edits against their own tasks.
The launch fits Polosukhin’s wider work on open and user-owned AI. He is a co-author of the 2017 paper “Attention Is All You Need”, which introduced the Transformer architecture, and now leads work at $NEAR AI, according to Fast Company. In a podcast interview summarised by Open Commons, he argued that releasing model weights alone does not make an AI system open source, and that real openness also requires access to training data, recipes and methodology.
Delta League tests a different route to improvement: crowdsourced, verified changes to a shared model, with results scored on data entrants cannot see. How far that approach can lift larger models remains to be shown, and the first results are small.
Sources