AI research project Marin opens 57 training checkpoints of the Mochi model.
Original headlineAI 研究项目 Marin 开放 Mochi 模型 57 个训练检查点
Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.
Open Athena said Nov. 17 that its Marin project’s Mochi 535B-A23B model has completed its main pretraining phase and released 57 training checkpoints, 10 of them from a long-context extension stage. The checkpoints have not undergone instruction tuning or safety post-training, so they are intended for scientific research and further training rather than direct deployment as assistants. The team also found data leakage in several commonly used pretraining benchmarks, which it said complicates fair comparisons of Mochi against other models. The release gives researchers intermediate snapshots of a 535B-parameter model to study training dynamics and build on. What to watch next is whether the leakage findings are quantified and whether instruction-tuned versions follow.
Muhabbit's AI wrote this summary in its own words from TechFlow's report and translated it from Chinese. It can get details wrong, so check the source.
Named in this story
Open Athena, Mochi 535B-A23B, Marin