Skip to content

AI Research Project Marin Releases 57 Training Checkpoints for the Mochi Model

Nov 19, 23:25

On November 19, Open Athena announced on November 17 that the Mochi 535B-A23B model in the Marin project has completed its main pre-training phase and released 57 training checkpoints, 10 of which are from the long-context extension phase.

Open Athena stated that these intermediate checkpoints have not yet undergone instruction fine-tuning or safety post-training, are intended for scientific research or further training, and should not be used as directly deployable assistants.

The team also noted that investigations found data leakage in some commonly used pre-training benchmarks, which affects the reportable evaluation results when fairly comparing Mochi with other models.

Source