NEAR Protocol co-founder Illia Polosukhin has launched Delta League, a platform that turns AI model improvement into a collaborative, crowdsourced process within the NEAR Protocol ecosystem. Polosukhin announced the initiative through his @ilblackdragon account. Instead of relying on a single development team, the platform allows outside contributors to propose changes to AI models, with every submission evaluated through verification procedures and private testing.

How Delta League Evaluates Model Improvements

Participants can submit several types of model modifications, including Low-Rank Adaptation (LoRA) files, sparse or dense weight differences, and complete model checkpoints. These submissions alter the numerical weights that determine how an AI model processes information and generates responses.

According to the Delta League website, the platform verifies submitted weights before running validation procedures and private tests. Each modification receives a score based on a hidden test set, enabling contributors to compare its performance against the original model without ever accessing the evaluation data itself. Keeping that evaluation data hidden prevents submissions from being tuned directly to the questions being scored, a known weakness when test sets are publicly available.

The system also preserves the lineage of every verified model version. Contributors can use previously submitted improvements as the starting point for further modifications, and each new version remains linked to the model it changes.

First Competition Targets Qwen3.5 0.8B

The inaugural competition is built around Qwen3.5 0.8B, a relatively small open model from Alibaba's Qwen family, with the 0.8B designation indicating roughly 0.8 billion parameters. Five verified modifications have so far increased its reported score from 18.6% to 19%.

Although the improvement is limited, the initial results provide an early demonstration of how Delta League's verification and evaluation process performs on an active model.

Architecture Changes and Business Testing on the Roadmap

The platform's next development phase is expected to expand beyond modifications to model weights. Support for changes to model architecture is planned for the coming days, potentially allowing participants to propose more fundamental adjustments to how models are structured.

Delta League also plans to introduce custom testing capabilities for businesses. These tests would allow companies to evaluate submitted modifications against tasks and requirements specific to their own operations.

Adding architecture-level changes and business-defined evaluations could broaden the types of contributions accepted by the competition. However, both capabilities were described as upcoming features rather than functions already available at launch.

Polosukhin's Broader Vision for Open AI

The initiative reflects Polosukhin's wider interest in open and user-owned artificial intelligence. He co-authored the 2017 research paper "Attention Is All You Need," which introduced the Transformer architecture that underpins most modern large language models, and now leads work at NEAR AI, according to Fast Company.

In a podcast interview summarized by Open Commons, Polosukhin argued that making model weights publicly available is not sufficient to qualify an AI system as open source. He also highlighted the importance of access to training data, training recipes and development methodology.

Delta League puts that philosophy into practice by treating model development as a shared improvement process: contributors submit changes, have them evaluated against private data, and make verified versions available for further development.

Whether this crowdsourced approach can deliver meaningful improvements to larger and more complex AI models remains uncertain. The initial Qwen3.5 0.8B results show a measurable but modest gain, while broader performance will need to be demonstrated through future submissions. Near-term markers include the arrival of the promised architecture-editing support and any further movement in the Qwen3.5 0.8B score past 19%.

Source: BSC News