Run the same task locally and in cloud.

Video caption

Run the same task locally and in cloud. The average can hide where each route fails. Local for bounded wins; hosted for measured hard wins. #EricFieldNotes

Full written post / accessibility read

Teams love a clip where the local model answers before the API spinner moves. But was it the same task, output quality and concurrency? A faster wrong answer is not a win, and a network delay may still dominate on a tiny decision.

Use routine labels, rare wording, long context and disconnected-network cases. Add a peak of simultaneous requests. A local model may dominate one slice while the cloud path dominates another. Without slices, the A/B answer tells you almost nothing about routing.

Start the timer at the user's request and stop when an independent rubric accepts the result. Log first token, full completion, failed calls, human correction, data boundary and total cost. Pin the exact local box and hosted model version so the result can be repeated.

Use each path only in the classes where it meets quality, latency, privacy and recovery. Preserve an explicit exception route. Do this because the right architecture is a measured allocation of work, not a contest between two logos.

#EricFieldNotes

Four-beat scene transcript

1. Run the same task locally and in cloud.

Teams love a clip where the local model answers before the API spinner moves. But was it the same task, output quality and concurrency? A faster wrong answer is not a win, and a network delay may still dominate on a tiny decision.

Visual: Do not compare a warm toy prompt with a remote hard case.

2. Make the hard cases visible.

Use routine labels, rare wording, long context and disconnected-network cases. Add a peak of simultaneous requests. A local model may dominate one slice while the cloud path dominates another. Without slices, the A/B answer tells you almost nothing about routing.

Visual: The average can hide where each route fails.

3. Time to the accepted answer.

Start the timer at the user's request and stop when an independent rubric accepts the result. Log first token, full completion, failed calls, human correction, data boundary and total cost. Pin the exact local box and hosted model version so the result can be repeated.

Visual: Include queue, retries, rescue and policy.

4. Split by proven fit.

Use each path only in the classes where it meets quality, latency, privacy and recovery. Preserve an explicit exception route. Do this because the right architecture is a measured allocation of work, not a contest between two logos.

Visual: Local for bounded wins; hosted for measured hard wins.

Research and claim limits

• Moonshot AI: Kimi K3 official model card (S128)
• Apple: current Mac mini technical specifications (S129)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.