According to reports, the controversial large model SubQ, which previously claimed to reduce computational consumption by a thousandfold, has released a technical report for its 1.1 Small (small-parameter) version. The development company Subquadratic collaborated with third-party evaluator Appen to conduct a tripartite assessment, claiming the model achieved a 98% retrieval accuracy at a maximum sequence length of 12 million tokens and performed nearly on par with state-of-the-art models in real-world programming benchmarks.

The technical report reveals that the model was not trained from scratch but rather derived by replacing the attention computation mechanism on top of an open-source frontier model, followed by incremental training on 1 trillion tokens. Currently, the model's core parameters are not publicly available for download, nor is there a publicly accessible API interface for general use.