Choose a local model from the workload backward_

Local inference conversations often start with parameter count. The practical constraints are memory, context, concurrency, latency, quantization quality, and the exact task the model must perform.

A model that barely loads can be less useful than a smaller model that serves two developers reliably and leaves room for embeddings or context.

What changes in practice_

Benchmark representative prompts, record time-to-first-token and memory, cap context, and route complex work to a fallback rather than pretending one model fits everything.

• Measure on your own hardware.
• Separate interactive and batch workloads.
• Define a fallback before production.

Our take_

The durable advantage is not adopting the newest tool first. It is building the identity, state, evidence, and operating boundaries that let a real team own the system after launch.