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, also, The idea is not to retrain Sora locally or build a full data center at home. The model would already be fully trained by OpenAI, and the company could simply provide two local options: a compressed/lightweight version for users with more modest hardware, and a full version for users whose hardware can handle it.
The local version would primarily be for one user, not for serving a huge number of users. The goal is also to preserve model continuity through checkpoints for each existing Snapshot/Alias, if such checkpoints actually exist internally. Each checkpoint could receive subsequent official updates—improvements, additions, further training, expanded creativity, etc.—while preserving earlier states for local experimentation and comparison.
For example, users could test how the model evolved across versions such as 2.1, 2.2, 2.3, 2.5, 3.0, and later releases. This would allow researchers and users to compare capabilities over time without requiring OpenAI to retrain the model from scratch for local use.
Of course, this assumes that the corresponding checkpoints exist and that OpenAI has actually maintained separate Snapshot/Alias states. It is simply a proposal for making an already-trained model available locally, not a proposal to recreate Sora’s massive training infrastructure at home.
