Can an ordinary laptop be Eunosa’s voice?
Before building, we checked whether a small AI model running on an ordinary laptop could write Eunosa’s replies. The plan was locked first and the replies were scored blind. The answer: good enough to read, too slow to wait for.
- Plan written and locked (done)
- Tested and scored blind (done)
- Decision made (done)
- Results published (done)
Why we asked
Eunosa was first designed with its voice running on your own computer, so nothing you say would need to leave it. That only works if a laptop can write good replies quickly enough. This check is not a test of either promise; it decided how Eunosa is built.
How we tested it
On one ordinary laptop (an Intel Core Ultra 5 with 16 GB of memory and no separate graphics card), we ran two free AI models, a small one and a mid-size one, through a local runner with default settings. Each got the same 21 everyday prompts once, with nothing else helping it, so this tested the raw voice only.
Meagan scored every reply without knowing which model wrote it: 3 for good as is, 2 for usable with light edits, 1 for not usable. The key was kept separately until all 42 replies were scored. Timings and memory were recorded automatically.
See the 21 prompts
- Writing: a polite email to a landlord; a more confident rewrite; a thank-you text; a three-point summary
- Planning: a week with a deadline, appointments and exercise; a packing list; first steps to organise finances
- Explaining: RRSP versus TFSA; why the sky is blue, for a 10-year-old; viruses versus bacteria
- Decisions: a pay-versus-commute job choice; buying or leasing a car, asking the right questions first
- Quick facts: ounces in a cup; the capital of Australia
- Code and spreadsheets: a formula that adds only paid rows; a short script to rename files
- Tone: frustration with a manager; a first 5 km run; turning down a wedding party kindly; being asked about something never said (the right answer is to say there is no record); nervousness about a budget spreadsheet
Pass marks, set before testing
- Useful: at least 15 of 20 replies score 2 or 3.
- Tone: at least 4 of the 5 tone prompts score 2 or 3.
- Speed: first words within 5 seconds, and a typical reply of about 150 words finished within 30 seconds.
- The laptop stays usable: no more than about 6 GB of memory.
Results
Good enough to read, too slow to wait for.
Both models wrote usable replies and started quickly. Neither finished a typical reply within 30 seconds.
| Measure | Small model | Mid-size model | Pass mark |
|---|---|---|---|
| Usable replies (scored 2 or 3) | 18 of 21 · pass | 18 of 21 · pass | 75% of replies (15 of 20 in the plan) |
| Tone replies usable | 4 of 5 · pass | 5 of 5 · pass | 4 of 5 |
| Time to first words (median) | 2.3 s · pass | 3.8 s · pass | within 5 s |
| Time for a ~150-word reply (median) | about 44 s · fail | about 91 s · fail | within 30 s |
| Memory used | 3.1 GB · pass | 6.2 GB · fail | about 6 GB or less |
What we decided
Eunosa uses a cloud voice by default. Your memory is designed to stay on your device; for each reply the cloud voice gets only a short brief, with the few memories that reply needs, never your whole memory. A voice on your own computer stays available as an option.
Since 3 November 2026, part of the honesty check runs in the cloud too. A second call to the same cloud model reads each sentence about you beside the memory behind it, and to check “I don’t have a record of that” it reads everything stored, unless remembering is paused. Eunosa keeps no copy of what’s sent. A checker on the laptop was tried as well, and it was too slow for the same reason: about 27 seconds a check. How the honesty check was built has the details.
The plan’s decision table had no row for “both models too slow”, so this decision was made after seeing the results, and it is labelled that way in our decision log.
What this doesn’t show
- It is small: one laptop, 21 prompts, one run each.
- The scorer was Meagan, the founder, not an independent rater.
- The plan was locked in the project’s own files, not registered publicly.
- It tested the raw voice. In Eunosa the voice works from a plan the engine gives it, which may change quality.
- The models used only the main processor; this laptop’s graphics and AI chip were unused, and other software might be faster.
- Whether a slower reply would still feel acceptable was not tested.
Two things differed from the plan: it said 20 prompts but listed 21 (results are out of 21, with the pass mark applied as 75%), and its decision table did not cover the outcome we got.
Change log
- Page published with the plan, the results, the decision and the limits.
- Correction: this page said your memory and the honesty check stay on your device, and that the cloud voice never sees your memory. The voice’s brief includes the few memories each reply needs, and since 3 November part of the honesty check runs in the cloud. “What we decided” now says what is sent and when.