The honest part other labs won’t admit: Over LAN we saw near-0% loss vs running solo. But sharding isn’t about speed, an M1 Mac + M3 Mac won’t beat a single M3. You save storage. You pay in latency.

Early split-inference was rough; ~2-4 tokens/sec. The handoff worked, but orchestration overhead ate everything: not enough model execution stayed hot. The breakthrough: moving the heavy path into…