Storage efficiency for Kimi K3 model checkpoints, specifically supporting quantized or compressed formats.

### What can't you do today

The current implementation is extremely impressive, especially the ability to stream weights from disk and keep RAM usage low.

However, the storage requirement is still a major barrier for many users. The official checkpoint is around 1.56 TB, which makes it impossible to experiment on consumer laptops and desktops with 1 TB SSDs, even though they may have enough RAM to run the streaming implementation.

As a result, many users who could otherwise run the project are unable to try it locally.

### What you would like

Would it be possible to support a storage-efficient version of Kimi K3 in the future?

I'm thinking of something like:
- support for quantized checkpoints
- support for compressed model formats
- compatibility with future community-compressed or distilled Kimi K3 variants

Even if generation becomes slower or quality decreases somewhat, having a version that fits within roughly 50-100 GB would make the project accessible to a much wider audience.

I understand this may be outside the scope of the repository or depend on future community work, but I wanted to ask whether this is something you've considered.

Thanks for creating such an amazing educational project!

### Checked the roadmap

- [x] I read docs/ROADMAP.md and this is not already listed