We publish only numbers we measure ourselves, and we say how to reproduce them. Where a figure has not been measured under a documented method yet, we say that instead of estimating.
This compares deployment models, not specific vendors — and it is a cost and compute comparison, not a capability benchmark. Large hosted models remain more capable on many tasks than a laptop-scale local model.
| Metric | NC AI (local CPU) | Hosted AI subscription | Hosted AI, pay-per-token | Self-hosted open-weight model |
|---|---|---|---|---|
| Recurring cost | None | Typically a monthly seat fee | Scales with usage | No licence fee; GPU hardware required |
| Hardware required | Any CPU — no GPU | None (vendor-managed) | None (vendor-managed) | Datacenter-class GPU |
| Where data goes | Stays on your machine | Sent to the provider | Sent to the provider | Stays on your infrastructure |
| Works offline | Yes | No | No | Yes, after download |
| Credentials needed | None | Account + payment method | API key + payment method | None |
| Runtime footprint | 570KB binary | N/A — hosted | N/A — hosted | Multi-GB model files |
Figures for NC AI are measured on a consumer laptop CPU (ARM64) using the runtime's own benchmark command. Reproduce them yourself: nc ai benchmark. Characteristics listed for other deployment models are general properties of those models, not measurements of any particular product.
We haven't yet published a formal Environment / Hardware / Workload / Reproduction-instructions writeup for every number on this site. As that documentation is finished, it will be linked from here rather than replacing the summary numbers above.