# 512 GB Mac evaluation record

An unfilled worksheet for your own measurements. Copy one run record for each model, quantization, context length and repeat. Keep your scripts, settings and results before rental access ends.

AI KIZAI rents dedicated remote access to an M3 Ultra with 512 GB memory and a 4 TB SSD, hosted in Kobe, Japan. The worksheet is free; Mac rental is paid and subject to availability. M3 measurements do not predict M5 performance. Complete the M5 record only after measuring M5 hardware.

## Evaluation plan

- Evaluation date / time / timezone: 
- Task or dataset identifier: 
- Quality criterion and pass threshold (set before testing): 
- Model / quantization / context variants to compare: 
- Number of repeats per variant: 
- Scripts and results saved to: 

## Run record (copy for each run)

| Field | M3 Ultra measurement | M5 remeasurement |
| --- | --- | --- |
| Measured date / time / timezone |  |  |
| Machine / chip / installed memory / SSD |  |  |
| macOS version |  |  |
| Model artifact / repository |  |  |
| Model revision / file checksum |  |  |
| Quantization / format / file size and units |  |  |
| Runtime / package version / commit |  |  |
| Command / script revision / runtime settings |  |  |
| Configured context length |  |  |
| Prompt or input identifier |  |  |
| Actual prompt tokens ingested |  |  |
| Actual tokens generated |  |  |
| Temperature / seed / sampling settings |  |  |
| Load succeeded? (yes/no) / error |  |  |
| Model load time (seconds) |  |  |
| Peak memory value |  |  |
| Memory measurement tool / method / units (GB or GiB) |  |  |
| Swap observed / measurement tool |  |  |
| Prompt-processing time (seconds) / method |  |  |
| Prompt-processing rate (tokens/second) / method |  |  |
| Generation time (seconds) / method |  |  |
| Generation rate (tokens/second) / method |  |  |
| Time to first token (if measured) / method |  |  |
| Quality criterion / threshold |  |  |
| Observed output / result file |  |  |
| Quality result (pass/fail/unresolved) |  |  |
| Workflow / integration result |  |  |
| Repeat number / warm or cold state |  |  |
| Other workload / OS state / caveats |  |  |

## Decision after measurement

- Which variant met the quality criterion?: 
- Which variants loaded at the intended context length?: 
- Observed limitations / unresolved issues: 
- What the M3 result supports: 
- What still needs M5 hardware testing: 
- M5 remeasurement date / matching configuration / differences: 
- Final decision based on both measured records: 

Record tool-reported units and timing boundaries. Peak process memory, system memory and GPU allocation are different measurements. Use the same inputs and settings for repeats, and record software or configuration changes. Leave values blank when unmeasured; do not extrapolate M5 values from M3.

- Evaluation guide: https://ai-kizai.com/guides/m5-ultra-512gb
- Existing M3 measurements and conditions (July 2026): https://ai-kizai.com/benchmarks/512gb