labzone-bench is a public, script-scored benchmark of how large language models — local and remote — work as coding agents on consumer mini-PCs and laptops. It includes iGPUs, reasonable quantization and thermal limits. It is an internal LabZone Labs measurement, not client work.
What we measure
Nine tasks (analysis, planning, code) run three times each, 243 checks in total, scored by the pi harness. We publish scores, per-check results, timings, token counts, model IDs, quantization, sampling settings, backends, hardware, and written findings. We withhold the task prompts, fixtures, answer keys, and the specifics of traps built into the tasks, so that knowing a trap exists is a hint and knowing its content is the answer.
Hardware
As described on the bench site: a faster “Strix Halo” machine and a realistic mini-PC tier (HX 370 / Strix Point). Scores transfer closely across those machines; time does not. Local runs typically use a 64k context and Q4-class GGUF.
Current results
See bench.labzone.tech for the live table. Method: methodology. What the numbers can and cannot tell you: limitations. Per-task detail: results.