FIELD NOTE / X
“It loads” is not “it serves.”
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved X edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
“It loads” is not “it serves.”
Video caption
“It loads” is not “it serves.” Offload, cold start and queueing can dominate. Keep the placement that meets the whole contract. #EricFieldNotes
Full written post / accessibility read
When someone says a frontier-scale model runs locally, ask what they measured. A process accepting weights, or emitting one token, does not prove latency, concurrency, quality or recovery at the workload you care about.
A demo may be warm and single-user. Now put two real users on it and include the long prompt, tool result and follow-up. Weight movement, cache pressure and contention may turn the same model into a poor service even if each isolated request eventually completes.
Use a fixed set of representative tasks. Record time to first token, sustained output, p ninety-five completion under stated concurrency, memory pressure, failed requests and independently judged answer quality. Compare local, hosted and smaller-model routes on the same cases.
Use local inference where privacy and accepted performance justify it. Use a hosted cluster where the full model and shared serving envelope matter. Do this because a loaded model is a software state; a service is a repeatable customer outcome.
#EricFieldNotes
Four-beat scene transcript
1. “It loads” is not “it serves.”
When someone says a frontier-scale model runs locally, ask what they measured. A process accepting weights, or emitting one token, does not prove latency, concurrency, quality or recovery at the workload you care about.
Visual: A launch banner says nothing about useful answers.
2. The first user finds the bottleneck.
A demo may be warm and single-user. Now put two real users on it and include the long prompt, tool result and follow-up. Weight movement, cache pressure and contention may turn the same model into a poor service even if each isolated request eventually completes.
Visual: Offload, cold start and queueing can dominate.
3. Write the service contract.
Use a fixed set of representative tasks. Record time to first token, sustained output, p ninety-five completion under stated concurrency, memory pressure, failed requests and independently judged answer quality. Compare local, hosted and smaller-model routes on the same cases.
Visual: Cold and warm delay, output rate, quality and recovery.
4. Benchmark the product, not the boot.
Use local inference where privacy and accepted performance justify it. Use a hosted cluster where the full model and shared serving envelope matter. Do this because a loaded model is a software state; a service is a repeatable customer outcome.
Visual: Keep the placement that meets the whole contract.
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.