Magic releases HashHop for long-context evaluation

A public tool for testing long context

Magic introduced HashHop in its August 29, 2024 research update and linked the implementation. The benchmark uses random hash associations and chains of retrieval steps to test long-context architectures while reducing semantic clues that can make a retrieval task easier. [1]

The public version 1.0.0 release was published that day, and the implementation carries an MIT license. The dated artifact is the evaluation tool. It does not establish public weights or general access to the internal LTM models described in the same company update, nor independent validation of their performance. [2][3][1]