NEAR Protocol introduced Delta League on November 5 as an open competition for improving AI models through direct changes to their weights. Its announcement describes submissions being scored on a hidden test set, with verified versions available for others to build on, and identifies NEAR AI as powering the initiative.
The NEAR announcement describes a competition, rather than the release of one new base model. The concrete developer task is to produce a modified version of an existing model and submit that change for evaluation.
Delta League’s published workflow starts with a registered parent model version. It accepts LoRA updates, sparse or dense weight differences, and full checkpoints. These are different artifact formats for supplying model changes; the site says submissions retain their model lineage, linking a change to its parent version.
The participation instructions separate public validation data from private tests. They direct participants to use a pinned evaluation configuration and compare the submitted version with its original model under that same configuration. A higher score would therefore need to be assessed within the relevant evaluation, rather than treated as a universal improvement across tasks.
The documentation also describes a research pull request containing the hypothesis, reproducible code and validation findings, with its URL attached to a submission. The command-line tool does not create that pull request automatically, and the page describes a source-free submission option.
For developers following NEAR’s AI work, the practical change is a documented route from a parent checkpoint to a weight-edit artifact and a scored submission. Evaluation uses a participant’s quota, so the instructions require checking it before scheduling an evaluation.
The developer-facing artifact is a versioned model change and its evaluation record. NEAR’s post identifies the initiative as powered by NEAR AI; it does not describe changes to NEAR blockchain consensus or transaction execution.