Announcement
Open Athena announces 57 Mochi checkpoints—and a benchmark-contamination catch
Open Athena says the release includes 10 long-context checkpoints, but these aren’t ready-to-deploy assistants.
TLDR
In its announcement, Open Athena says 57 Marin “Mochi” 535B-A23B checkpoints span training across 18.2 trillion tokens, including 10 from long-context extension. It also reports substantial benchmark leakage its decontamination process missed, limiting fair model comparisons. The team warns these checkpoints have no instruction tuning or safety post-training: research material, not a finished assistant.
Combined views
10.1K
10 Sources, first seen ago

Comments