Operator-owned: hardware and Docker on the sandbox host. 8e verified: Winnow-12B Q8 needs about 15 GiB peak at 64k context (12.7 GB file), openjev Q4 needs a 24 GB card, FP8 needs 80 GB; the razorback16/openjev Docker image serves the same wire API (NVIDIA 24 GB or Apple 16 GB). The card in the box is UNVERIFIED - the repo knows it only as the hub on :8025. Route direct, not through the LiteLLM gateway (version unverified, 90 s ceiling measured). When the endpoint answers, 8a declares the local part with its BaseUrl.
plans/spway/8e/report.md
docs/memory/reference_litellm_gateway_capacity.md