See also "pain points" from Hybrid AI Requirements presentation on 2024-04-04. Marked as draft since we may still want to revise these.
- Saving space
- Minimizing download latency
- Sharing/reusing large models
- Across sites
- Same resources at different URLs
- Same resources in different formats
- Want to expose "built-in" models
- Generalizing models
- Categories (e.g. "en-jp text translation")
- Versions (e.g. ">= 1.3.*")
- Need solution that can handle adapters
- Applications sharing a base model
- Foundational models are large
- Adapters are small
- Supporting model updates
Note: Most of the above currently addresses the problems of "model management", not distributing models between client and server (e.g. "hybrid AI").