A loaded model has six unanswered questions.

Video caption

A loaded model has six unanswered questions.

Cold start, long context, contention, errors and recovery.

Latency, throughput, concurrency, quality, recovery, cost.

Keep the smaller model when it passes; escalate when it does not.

#EricFieldNotes

Full written post / accessibility read

A model screenshot proves the files were found. It does not tell you when the first token arrives, how fast the answer finishes, how many users fit, whether the answer is good, what a failure costs or who operates it.

Imagine a local assistant used by a support team. The first warm question is fine. At shift change, several long records arrive at once. The queue grows, one request fails, and no one knows whether the retry reuses context or starts over. The initial demo cannot price that experience.

For the exact model build and hardware, measure cold and warm first token; completed responses per minute at stated concurrency; accepted answer quality; restart and retry behavior; and full operating cost. Keep task difficulty and input sizes visible.

Use a compact local model for bounded private work that clears the six-cell contract. Escalate difficult or high-concurrency work to a stronger service. Do this because local ownership is useful only if users receive a timely accepted result.

#EricFieldNotes

Four-beat scene transcript

1. A loaded model has six unanswered questions.

A model screenshot proves the files were found. It does not tell you when the first token arrives, how fast the answer finishes, how many users fit, whether the answer is good, what a failure costs or who operates it.

Visual: The screenshot is the smallest part of deployment.

2. One happy prompt hides five cliffs.

Imagine a local assistant used by a support team. The first warm question is fine. At shift change, several long records arrive at once. The queue grows, one request fails, and no one knows whether the retry reuses context or starts over. The initial demo cannot price that experience.

Visual: Cold start, long context, contention, errors and recovery.

3. Use a six-cell acceptance card.

For the exact model build and hardware, measure cold and warm first token; completed responses per minute at stated concurrency; accepted answer quality; restart and retry behavior; and full operating cost. Keep task difficulty and input sizes visible.

Visual: Latency, throughput, concurrency, quality, recovery, cost.

4. Only call it local when it serves.

Use a compact local model for bounded private work that clears the six-cell contract. Escalate difficult or high-concurrency work to a stronger service. Do this because local ownership is useful only if users receive a timely accepted result.

Visual: Keep the smaller model when it passes; escalate when it does not.

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.