Open Sora 2 Local Checkpoints for Model Continuity
I propose that, if Sora 2 models are ever released for local use, their training checkpoints should also be made openly available.
For each official Snapshot and Alias, users could access checkpoints representing different stages of the model’s development, including checkpoints from before and after any potential capability reduction, training adjustment, or other model modification — if such changes actually occurred.
The exact timing and nature of these changes should not be assumed; the purpose is to preserve the model’s actual development history rather than speculate about it.
Each checkpoint could then support future official versions (Improvements, More training) for each update, and more official updates ( 2.1 2.2 2.3 2.5 3.0), allowing users to maintain, compare, restore, or build upon different stages of the same model family.
This would provide long-term model continuity, reproducibility, transparency, preservation, and greater control for local users, especially after cloud versions are discontinued or replaced.
