Local does not automatically mean faster.

Video caption

Local does not automatically mean faster.

One request is a poor service benchmark.

Same cases, same deliverable, named hardware and builds.

Use the route that meets each case's acceptance budget.

#EricFieldNotes

Full written post / accessibility read

A small Qwen on local hardware avoids a remote round trip. A hosted Haiku or typed Jev route may avoid local queuing or token generation work. None of those facts determines the winner on your actual service at your actual concurrency.

A local test can look instant when the model is warm and alone. Add simultaneous users, a cold restart and a long explanation, and the tail may change. A remote route has its own network and provider limits. These are questions for a matched workload, not a tweet about parameter count.

Measure warm and cold p fifty and p ninety-five time, batch concurrency, abstentions, retries and validated outcome. Include local memory and operator cost, remote network wait, and any extra writer or reviewer. Fix model versions and record settings so the result can be repeated.

Keep simple work on the cheapest path that passes the quality and tail-latency gate, and escalate hard cases. Do this because local versus cloud is an operating condition; accepted work is the product.

#EricFieldNotes

Four-beat scene transcript

1. Local does not automatically mean faster.

A small Qwen on local hardware avoids a remote round trip. A hosted Haiku or typed Jev route may avoid local queuing or token generation work. None of those facts determines the winner on your actual service at your actual concurrency.

Visual: Network latency is only one part of a decision path.

2. A warm demo hides the queue.

A local test can look instant when the model is warm and alone. Add simultaneous users, a cold restart and a long explanation, and the tail may change. A remote route has its own network and provider limits. These are questions for a matched workload, not a tweet about parameter count.

Visual: One request is a poor service benchmark.

3. Benchmark the operational envelope.

Measure warm and cold p fifty and p ninety-five time, batch concurrency, abstentions, retries and validated outcome. Include local memory and operator cost, remote network wait, and any extra writer or reviewer. Fix model versions and record settings so the result can be repeated.

Visual: Same cases, same deliverable, named hardware and builds.

4. Deploy by slice, not ideology.

Keep simple work on the cheapest path that passes the quality and tail-latency gate, and escalate hard cases. Do this because local versus cloud is an operating condition; accepted work is the product.

Visual: Use the route that meets each case's acceptance budget.

Research and claim limits

• TypeSafe AI: Introducing System One Models and Jev (S97)
• Anthropic: Models overview (S104)
• Qwen: Qwen3 model family (S105)

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