Existing local measurements: what they establish
Measurement companion · September 13, 2026. No new runs were performed for this refresh. These are observations from particular configurations, not a controlled comparison of Colibri versions, CPU versus GPU, cloud efficiency or completed-task energy.
| Retained test | Observation | Important boundary |
|---|---|---|
| Completed GLM-5.2 weather workflow, September 11 | 26:02.24 timed turn; two model rounds; one successful weather tool call. Observation retrieval took 224 ms. | Activation/setup excluded. Answer contract was explicitly bounded. Earlier attempts failed or timed out. This is one completed task, not ordinary request latency or a hardware comparison. |
| GLM-5.2 process-start/reuse diagnostic, September 11 | First 458-token request: 476.622 seconds; identical repetition: 6.756 seconds. | Both deliberately limited to one output token, finish_reason:length. They were diagnostics, not completed-answer acceptance tests. Initial request had 456 new tokens; repetition reused the whole prefix. OS/storage caches were not cleared. |
| CPU-only GLM-5.2 candidate, September 12 | Short readiness reply: 22.122 seconds. New-prefix probe timed out after 300.809 seconds without generated output. | CPU execution, RAM budget and pinning changed together. Earlier GPU diagnostic ran longer than this deadline. The timeout is a censored observation, not proof that one arrangement won. |
| Reduced pin-budget experiment, September 11 | New-prefix diagnostic: 537.782 seconds; identical repetition: 7.112 seconds. | Matched baseline comparisons did not finish qualification because resource checks refused continued operation. No isolated speedup can be calculated. |
The weather reply's required station, measured units, source and observation-age fields were checked. No claim is made that the data were current at final delivery: age was labeled at retrieval. The final answer was an observation, not a forecast.
Existing instrumentation also showed why metric definitions matter. A model profile could omit initial processing while returning a near-zero window; sampled process read transfers could include cached I/O. Neither proves a zero-duration prefill or physical-disk throughput. Separate wall-clock, phase, cache and device measurements are necessary before assigning a cause.
No trustworthy whole-system electricity comparison, cloud baseline, avoided-cloud-usage measurement, adoption study or data-center-capacity effect was established by these tests. The associated article makes none of those outcome claims.