Deltafin makes local MoE serving an I/O tradeoff

The Deltafin repo exposes an OpenAI-compatible server for Kimi K3 and offers a full local download or on-demand expert streaming. Its README lists roughly 1.7 TB for the full mode and about 215 GB for streaming. Why it matters: A local endpoint does not remove the operating constraint; it moves it into disk, network, and token latency. The streaming path lowers the storage commitment but can take 3+ minutes to fetch uncached experts.

Try this: Inspect the hardware, disk, network, and latency table before installing. Treat the maintainer's speed figures as a starting point for a reproducible test, not a capacity plan.