If you're not in a hurry, you don't need a frontier model to build with the anatomy skill. Here's a microwave built entirely on a MacBook Pro M5 Pro (64 GB) with Qwen 3.8 Flash Next, quantized to ~2.6 bits, served by Sushi and driven by Pi. One long end-to-end task with zero human input to the model: no prompts, no hints, not a line of the project's code from me. - 17 h 20 min run time - 76 agent sessions, 23/23 tasks done - 1,454 model responses, 1,676 tool calls - Decode 41.4 tok/s - Median TTFT 3.3 s (p90 14.5 s) - 66% MTP draft acceptance - 90% of prompt tokens served from the prefix cache - Peak prompt 110K tokens, only 0.35 GiB swapped - 15/16 hidden acceptance tests The point of this run is the approach, not the drawing. I don't think small local models fail at long tasks because they're dumb. They fail because nobody keeps them on rails. So that job moved out of the model and into a deterministic orchestrator: - Planning up front: the model turns the brief into requirements, task cards and a deterministic verify script. - Strict task order, one task at a time. - Gated acceptance: a task passes only if verify passes AND a fresh reviewer session says ACCEPT. - Replanning when reality diverges from the plan. - Context control: every session has a token budget; past it, the session is cut and a fresh one picks up. - Memory pressure: engine OOM errors are retried and don't count as failed attempts. - Retries with feedback until the task passes. - It doesn't give up: a blocked run resumes where it left off. Every role in that loop (planner, builder, reviewer, replanner, final reviewer) is the same small local model on the same laptop. The orchestrator is plain deterministic code with no LLM in it. No frontier model touches any phase of the run. The model only ever does one small, verifiable thing at a time. The orchestrator does the remembering, ordering and judging. My thesis: if wall-clock time doesn't matter, a small local model can do almost any task, provided the loop around it is deterministic and refuses to give up. This is a first attempt. The goal is a runner with enough guardrails that a model like Qwen 3.8, with a limited context window, can work for hours without drifting and still hold a minimum quality bar. This microwave is one example of that loop at work. The challenge: an interactive isometric cutaway of a countertop microwave that explains how it heats food. It shows the transformer, capacitor, diode, magnetron, waveguide, cavity, turntable and mesh door. Energy flows into the cavity and stops when you open the door. Power levels run as an on/off duty cycle, with water, a frozen block or an empty dish inside, and a live readout of power, time, energy and temperature. It ships dark and light pages, a geometry audit and screenshots. The project's own verify runs 52 checks. It also shows that a small local model can follow a demanding skill. Anatomy asks for many small, precise solid parts, four-tone shading, correct painter's order and calm, eased motion, and the skill was loaded in every session. The final build passes the skill's geometry audit and its line-end check, with 0 loose ends on both pages. Full run sheet, playable build and measurements: aartiles.github.io/local-mod…