Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the default ate the afternoon

Qwen 3.8's default reasoning mode can turn a local task into a twenty-one-minute wait; local control starts with inspectable defaults.

Published Updated Read 2 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-08-17

a 17 GB local model can now be handed a laptop-sized job, then quietly spend twenty-one minutes thinking about a pelican on a bicycle. Qwen 3.8’s default reasoning setting makes the cost knob somebody else’s first decision; at home, that is a bad place to leave it.

1. the default ate the afternoon

sources:

what happened:

Qwen’s Apache-2.0 27B vision model is available as a local model, and its documentation exposes reasoning_effort from low to xhigh. Simon Willison tested a 17 GB Q4 build on a MacBook Pro and found that its preserved default, xhigh, used 22,276 reasoning tokens and twenty-one minutes to produce a 3,223-token SVG. With reasoning off, the same prompt took 137 seconds. The slower run made a better picture; it also invented a small design brief around a request for a circle.

the collision:

Local weights do not automatically make a personal system yours. A default can still decide how much of your afternoon, battery, and attention gets spent before you see an answer. The useful capability here is not merely running the model at home; it is being able to inspect and set the trade between speed, cost, and fuss before the model turns a plain task into a private little epic.

question left open:

Which agent defaults deserve to be shown as decisions before they are allowed to become somebody’s invisible work style?

left on the table

  • Agentao has a sensible local-first governance design, but this is a newly posted architecture paper, not yet evidence that its permission and replay claims hold up in a real host environment.
  • Wild Static puts one shared memory in front of all visitors, but its public surface is too thin to establish what is retained, who can alter it, or what the experiment has taught anyone outside its own room.
  • craft coding makes a sharp case for using AI as reviewer rather than author. It stays out because the Qwen fieldnote already carries today’s question about preserving human judgment inside an AI-assisted workflow.
  • Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
  • Local LLM runtimes — pick local inference by custody, speed, privacy, cost, and failure mode
  • Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge
local-models personal-ai reasoning-controls