Dragonfly Adds Native Hugging Face and ModelScope Protocols

What happened

Dragonfly added native Hugging Face and ModelScope support to dfget via hf:// and modelscope://.

For platform teams, this removes a lot of custom model sync and cache glue.

You can pull models directly with built-in auth, revision pinning, and recursive fetches.

Why it matters

On inference clusters, model pull is usually the cold-start bottleneck and one of the biggest avoidable egress costs.

The traffic math is straightforward. A 130 GB model across 200 nodes is 26 TB if every node pulls from origin. With Dragonfly in front, origin should see roughly one full pull (~130 GB). That is the 99.5% claim.

It also reduces lock-in to internal mirror pipelines and custom fetch services that become permanent maintenance debt.

How it looks

dfget hf://deepseek-ai/DeepSeek-R1/model.safetensors -O /models/DeepSeek-R1/model.safetensors
dfget hf://owner/repo -O ./repo/ -r

ModelScope follows the same pattern with --ms-token and --ms-revision.

The key behavior is piece-level sharing. Seed peers can upload chunks before the full model download completes, so pulls become parallel instead of serialized.

Proven vs unproven

Proven:

• Native hf:// and modelscope:// support in dfget
• Cleaner integration path than custom mirror jobs
• Clear origin-traffic reduction math

Unproven (in public data so far):

• p50/p95 model-ready startup time at scale
• cross-AZ and cross-region behavior
• token handling guidance that avoids secrets in shell history
• failure behavior when origin is slow or rate-limited
• end-to-end security posture for private model pulls

What to do next

If you currently mirror Hugging Face or ModelScope into internal storage, run a controlled pilot and compare:

• cold-start time per pod
• origin egress
• east-west network load
• operational complexity (how much code/config you can delete)

If the numbers hold, delete the mirror pipeline.