60 client emails. Four models. Every answer.

Meow builds private AI for businesses that can't let their data out, and chooses each model by testing it first. This is one such test, in full: pick an email and compare what each model answered with the answer checked by hand.

Book a free 15‑minute call Run the two small models yourself

September salaries

Dear Ms Favre, Just to confirm: the September salaries were paid correctly and everyone is happy. No action needed. Thanks, Sarah Weber

	Checked by hand	0.6Bin a browser	1.7Bin a browser	8Bin a browser	32Bone server GPU	32B + rulessame GPU
Topics	Salary	Wrong: Salary, Accounting	Right: Salary	Right: Salary	Right: Salary	Right: Salary
Urgency	Low	Wrong: Normal	Right: Low	Right: Low	Right: Low	Right: Low
Deadline	None	Right: none	Wrong: current	Right: none	Right: none	Right: none

The right model is measured, not guessed

Added up over the 20 emails, in each language. 0.6B, 1.7B and 8B ran in a browser on an Apple M1 laptop; 32B ran on one server graphics card, the size of model a private deployment serves to a whole team. All four got exactly the same instructions.

AcademyWhether a model fits on a GPU AcademyHow to read model scores

Triage, measured: size aloneCorrect answers out of 20 emails per language, with the same instructions for every model. Times are medians: 32B on a server GPU, the others on an Apple M1 laptop.

English
	0.6B	1.7B	8B	32B
Deadline	19	14	18	15
Topics	11	12	10	11
Urgency	9	11	11	11
All three right	6	3	5	3
Time per emailMedian	1.3 s	2.4 s	6.6 s	1.5 s

French
	0.6B	1.7B	8B	32B
Deadline	12	12	20	19
Topics	1	2	9	8
Urgency	5	9	14	12
All three right	0	0	5	3
Time per emailMedian	1.5 s	3.0 s	6.7 s	1.7 s

German
	0.6B	1.7B	8B	32B
Deadline	13	12	20	18
Topics	3	7	4	6
Urgency	9	14	10	10
All three right	1	4	3	2
Time per emailMedian	1.5 s	3.1 s	7.4 s	1.6 s

The same 32B, with the rules written inThe instructions above name the fields and nothing more. Here they also state the rules the answers were checked against, word for word, as they were written before any model saw the emails.

English
	Same instructions	Rules written in
Deadline	15	20
Topics	11	10
Urgency	11	19
All three right	3	9

French
	Same instructions	Rules written in
Deadline	19	20
Topics	8	6
Urgency	12	19
All three right	3	6

German
	Same instructions	Rules written in
Deadline	18	20
Topics	6	7
Urgency	10	20
All three right	2	7

The rules, as added to the instructions

Rules: urgency is high when the client says it is urgent or a penalty or loss is imminent, low when the email is information only with nothing to do, and normal for any other request. The deadline is the date the client gives for the fiduciary to act, or none. The topics are every topic the email asks about.

What the test showed

A bigger model closes only part of the language gap. In French, 8B found every deadline, 20 of 20, and named the topics right 9 times, against once for 0.6B. In German it also found every deadline, but filed 19 of the 20 emails under "Other".
Bigger is not automatically better. With the same instructions, 32B got all three right less often than 8B in every language: it marked as urgent emails the checked answers rate as normal, and added topics the client didn't ask about.
Once a model is large enough, instructions matter more than size. With the rules written in, the same 32B judged urgency right in 19 or 20 emails of 20 and found every deadline, in all three languages. Spelled-out urgency rules had made the smallest model call almost every email urgent.
Topics are the hard part. Even with the rules, 32B picked exactly the right topics in only 6 to 10 emails of 20, mostly by adding a topic the email only mentions. That is what examples from your own emails are for.

Read the sample report

With the rules written in, the same 32B model went from 3 to 9 emails of 20 with all three right in English. What remains is judgement on topics: the next step is examples from your own emails, checked with the same test.

Book a free 15‑minute call

Agents that ask before they act

Stackive In daily use

Stackive is the platform built in-house to run the company: email, invoicing and deployments go through it, each in a service of its own, with one agent layer that works across them.

Its agents read what a task needs and prepare the work. Anything that sends an email, issues an invoice or deploys code waits for a person on the team to approve it, and every step is recorded. The rough edges show up here first, before the same kind of agents are set up for you.

Every entry names who acted, an agent or a person, and nothing leaves until a person has said yes.

A morning in the log. An illustration with made-up entries.

08:57Email agentSorted 14 new emails into client foldersDone
09:02Email agentDrafted a reply to a supplier about a late deliveryWaiting for a person
09:05A personApproved the replyApproved
09:05Email agentSent the replySent
10:31Billing agentPrepared an invoice for 12 hours of workWaiting for a person
10:40A personSent it back: wrong hourly rateSent back
10:42Billing agentCorrected the rate and prepared it againWaiting for a person
11:15Delivery agentPrepared version 4.2 for releaseWaiting for a person

Accessible software

Voice search that stays on the device
Local voice assistant Prototype, 2026
A French voice assistant for searching and playing an audiobook catalogue hands-free, running entirely in the browser: Whisper speech recognition, a language model that turns spoken requests into catalogue searches, and speech synthesis, all on the listener's device, with no cloud speech service.
Every model and prompt change is scored against a hand-checked corpus of real voice requests, so improvements are measured rather than assumed. Once loaded, the interface reopens without a connection.
Runs on
The listener's device
Models
Whisper, a small language model, speech synthesis
Measured on
Real voice requests, checked by hand
On the listener's device
1. Spoken question
2. Transcribed, then searched
3. Spoken answer
An audiobook library in the browser
Web Player Web app, 2023
The Bibliothèque Braille Romande's catalogue as a web app: more than 14,000 audiobooks to search, listen to online or download, with a member login and help pages on braille and DAISY books.
Built for the keyboard and screen readers first: skip links to every part of the page, and the expected screen-reader behaviour written down as testable rules, checked by automated tests with Playwright and Axe. Its audiobook player and accessibility toolbar are published as open-source packages.
Runs in
Any web browser
Catalogue
More than 14,000 audiobooks
Checked with
Screen-reader tests and Axe audits
- Open the web player
Open-source building blocks
a11y-player and a11y-button npm packages, 2025
The web player's audiobook player and accessibility toolbar, published as React packages so any app can use them.
a11y-player plays DAISY audiobooks: fully keyboard-navigable, working with screen readers, with section navigation, playback speed and shareable bookmarks. a11y-button adds a toolbar where readers switch to high contrast, change text size and spacing, use a reading mask or dyslexia-friendly fonts, and keep their choices.
Licence
MIT
For
React apps
Used in
The Web Player
- a11y-player on GitHub
- a11y-button on GitHub
```
$ npm install a11y-player a11y-button

import { DaisyPlayer } from 'a11y-player';
import {
  AccessibilityProvider,
  AccessibilityToolbar,
} from 'a11y-button';

<AccessibilityProvider>
  <AccessibilityToolbar />
  <DaisyPlayer
    dirUrl={bookUrl}
    appUrl={appUrl}
  />
</AccessibilityProvider>
```
Existing apps, rebuilt for blind readers
BBR Player™ iOS / Android
The Bibliothèque Braille Romande's audiobook apps for iOS and Android started from an open-source codebase by the BSR in Lausanne. The code was refactored and the features its members asked for were added, following W3C accessibility guidelines and working with blind and partially sighted readers throughout.
After release, the apps received a distinction from the Swiss Federal Office of Culture.
Starting point
An open-source codebase
The work
Refactoring and new features
Recognised by
The Swiss Federal Office of Culture

60 client emails. Four models. Every answer.

The right model is measured, not guessed

What the test showed

Agents that ask before they act

Accessible software

Voice search that stays on the device

An audiobook library in the browser

Open-source building blocks

Existing apps, rebuilt for blind readers

Book a free 15‑minute call

What happens in 15 minutes

Your situation

What's possible

The next step

What should we take off your plate?