Local Model Workers — Lab Report Series

Measured experiments running agentic coding workers on local open-weight models, on one Apple workstation (M5 Max, 128 GB unified memory), against cloud yardsticks. Every number in these reports comes from a run on real tasks with recorded logs; nothing is estimated.

Papers

#	Date	Title	Status	PDF
1	2026-10-15	Can open-weight local models act as the workers in an agentic coding workflow?	lab report	paper.pdf
2	2026-10-16	Can a local coding setup run beside normal work? (serving stack, prompt shape, harness repair)	lab report	paper2.pdf
3	2026-10-17	Quantisation, serving and harness repair: what it takes to run a local coder beside real work	lab report — numbers pending daylight build verification	paper3.pdf

Each directory holds the LaTeX source, the compiled PDF and the shared bibliography.

Findings in one line each

1. Paper 1: No for real application fixes at night-1 state — the local models find the code only when the task names the file and function.
2. Paper 2: The bottleneck was the serving stack, not the model — fixing the prefix-cache path took a local model from unusable to ~2× Sonnet wall time on one-function fixes; the harness itself needs repair hooks (path repair) for local models to function at all.
3. Paper 3: Quant lift replicates across families (4-bit→8-bit: 4/6→5/6 on two families); a 30B-A3B at 8-bit delivers rank-1 quality at half the resident memory; SSD-paged 125B-class models complete real tasks at ~3 tok/s — paging, not the model, is the limit; the remaining failure classes (invented paths, dotted tool names) are model-side and partly harness-repairable.

Method notes

• Suite: paired A/B arms on replayed real fixes with known answers; buildless diff-grading during runs, build-verified verdicts in a separate daylight pass.
• Machine: one 128 GB M5 Max, one model server at a time, memory watchdog.
• Replication rule: no single-run verdicts.
• Instrumentation: PAI (open source), runs recorded in the reports.

Plans

• arXiv submission of the series once the daylight build-verification pass confirms paper 3's rankings.