Why it exists · ทำไมต้องมี
Language is the gate to the world's knowledge. We help you open it yourself, not have AI walk through it for you.
⌘T · Translate · แปล
Two answers, side by side.
Apple's Translation framework answers on your Mac. With the local model on, Buddy 1.0, Gemma 4 E2B fine-tuned for Thai finance, translates the same English beside it, and usually first: 0.43 s against Apple's 1.8 s for ten words.
Each answer is a card, and the words the two disagree on are underlined, so you see where the meaning is in doubt. Settings → Translate with picks Apple + Buddy, Apple only, or Buddy only. Thai → English is always Apple.
Sentences and answers come from the app's 76-row Thai test corpus, as each engine answered them, and the timings are the measured ones for a sentence that long. Neither is always right, which is why you see both.
The Fed will keep shrinking its balance sheet.
เฟดจะยังคงลดขนาดงบดุลของตนต่อไป
เฟดจะยังคงลดขนาดสมดุลของตนต่อไป
Underlined: where the two differ
Credit spreads widened to their highest level since March.
ส่วนต่างอัตราผลตอบแทนของตราสารหนี้กว้างขึ้นสู่ระดับสูงสุดนับตั้งแต่เดือนมีนาคม
ช่องว่างเครดิตกว้างขึ้นถึงระดับสูงสุดนับตั้งแต่เดือนมีนาคม
Underlined: where the two differ
Foreign investors were net sellers of Thai stocks for a fifth week.
นักลงทุนต่างชาติขายสุทธิหุ้นไทยเป็นสัปดาห์ที่ 5 ติดต่อกัน
นักลงทุนต่างชาติเป็นผู้ขายหุ้นไทยเป็นรายได้สุทธิเป็นสัปดาห์ที่ห้า
Underlined: where the two differ
The everyday sense, or the term of art
Most of what Apple misses is a single word read in its everyday sense. Across the 309 concepts in the corpus, Apple's answers carry 278 (90%) and Buddy's 299 (97%).
⌘G · Check · ตรวจ
Proofread your own English, gently.
Select a sentence you wrote, anywhere, and press ⌘G. Languagly lists each mistake, then the corrected sentence, then the same sentence in five tenses as a reference.
When your English is already right, it says so and changes nothing. Getting that right took measuring: an early prompt invented a mistake in every correct sentence it saw.
The sentences, fixes and other ways to say it are Buddy 1.0's answers from the 200-sentence test corpus. The labels, rules, verb forms and tenses are what the app shows for them.
She work at the bank.
workworks
Singular and plural disagree here.
“The fund holds” but “the funds hold”. The -s moves to the verb when the subject is singular — the opposite of what it does on a noun.
She works at the bank.
CopyShe works at the bank.
She worked at the bank.
She has worked at the bank.
She is working at the bank.
She will work at the bank.
We plan to reducing exposure.
reducingreduce
The verb needs a different form.
reduce
After to, and after do / does / did.reduced
Stands alone as the past.reduced
After have / has / had.reducing
After am / is / are / was / were.We plan to reduce exposure.
CopyWe plan to reduce exposure.
We planned to reduce exposure.
We have planned to reduce exposure.
We are planning to reduce exposure.
We will plan to reduce exposure.
We plan to reduce our exposure.
The goverment issues bonds.
govermentgovernment
A spelling near-miss of the same word.
The government issues bonds.
CopyYields rose sharply after the auction.
Looks good — no issues found.
Tap any word · แตะคำไหนก็ได้
Guess first, then check.
Select one word and press ⌘T: below the two answers come its dictionary entry, synonyms and antonyms, verb forms, and usage examples whose Thai stays hidden until you tap it.
When the dictionary only lists a word under its root, Languagly says where the meaning came from, and the local model writes the word's own Thai meaning. Gemma 4 E2B, Buddy's base, was right on 18 of 20 real words, in about 0.3 s each; Buddy has not been measured on this yet.
tightening
การคุมเข้ม
การบีบแน่น
Underlined: where the two differ
• ทำให้แน่นหนา
• ทำให้รัดกุม
• กลายเป็นแน่นหนา
• กลายเป็นรัดกุม
• เข้มงวด
• กวดขัน
tightener
tighten
After to, and after do / does / did.tightened
Stands alone as the past.tightened
After have / has / had.tightening
After am / is / are / was / were.Practice · ฝึก
Your mistakes become your lessons.
Every check is kept. Practice is built from your own corrections and spaced out so it sticks, and the Analytics window shows how your writing changes: clean checks by week, and each kind of mistake, earlier against later.
Analytics
589 translations · 133 checks · 67 active days
- agreement26 → 0
- missing word21 → 1
- capitalization8 → 7
- extra word7 → 2
- spelling7 → 2
The Fed will keep shrinking its _____ sheet.
Then practise what's due.
Practice takes the ten most overdue words at a time. A word you saved from a sentence comes back as that sentence with a gap; one without, as its meaning to match; with too few words saved, as a card to recall and grade, and the grade sets the next review: under a minute, a day, or more.
The sentence is from the test corpus. The counter is real: 17 words are due on the developer's Mac, and a session takes ten.
Glossary and wordbank
Save words as you meet them, with the sentence you met them in.
Spaced repetition
Drills from your own corrections, timed so you see each one just before you forget it.
A daily reminder
Once a day, how many saved words are due for review. An invitation, never a nag.
Ask
Ask about a mistake type or a word, and keep the answer.
Right-click → Translate
From the Services menu in any app, if you would rather click than press ⌘T.
One-file backup
Everything in one JSON file. Restoring merges, so it never deletes work done since.
Measured, not assumed · วัดจริง ไม่เดา
How the numbers got here.
A product that teaches has to be honest about how often it is right. Each step below was measured on this app, on a Mac, and carries the date of the release that shipped it, or the day it was measured.
- 2026-08-25v2.206.1 → 1.4 s
Gemma 4 E2B replaces Gemma 3 4B
A check takes 1.4 s instead of 6.1 s, and accuracy on the 200-sentence corpus goes from 86% to 91%.
- 2026-08-31v2.319 → 5 rules
Four checker rules deleted
Rules that refuse a bad correction. On 211 scored sentences four never fired once, and one refused a real fix.
- 2026-09-01v2.322,980 → 19 MB
The model rests when you do
It is released after ten idle minutes and loaded again when you need it.
- 2026-09-16v2.361.15 → 0.73 s
A shorter prompt
The model had stopped writing the per-mistake line the prompt asked for. The corrections come from comparing its rewrite with your sentence, so the line went.
- 2026-10-01v2.372.1 → 0.40 s
Gemma translates beside Apple
For ten words, Gemma's answer arrives in 0.40 s instead of 2.1 s, by translating your English directly rather than correcting Apple's draft. Apple's own takes 1.8 s.
- 2026-10-15re-run0.3 s
Re-measured: 116/126 caught, 8/74 false alarms
A check with the model warm, on the repacked model: 200 checks in 59 s, one load included. Accuracy 91%.
- 2026-10-19v2.393.55 → 2.05 GB
Repacked on your Mac
Audio removed, the image encoder cut to a stub, the word list cut to Thai and English: 3.00 → 2.15 GB in memory, a reload 1.91 → about 1.2 s, and every test answer byte-identical.
- 2026-10-21beta8 → 3 false alarms
Buddy 1.0 replaces Gemma 4 E2B
The same model tuned for Thai finance: Thai concepts 291 → 299 of 309, false alarms 8 → 3 of 74, about the same speed (0.43 s for ten words, against 0.40), the same 2.05 GB, and downloaded already packed. In the private beta build.
| model | caught | false alarms | accuracy | per check |
|---|---|---|---|---|
| Buddy 1.0 betaGemma 4 E2B, tuned for Thai finance | 116/126 | 3/74 | 94% | 0.3 s |
| Gemma 4 E2B 2.39 | 116/126 | 8/74 | 91% | 0.3 s |
| Gemma 3 4B retired | 109/126 | 12/74 | 86% | 6.1 s |
| Qwen 3.5 4B retired | 92/126 | 1/74 | 83% | 2.0 s |
Why the most precise model lost. Qwen altered only one correct sentence, but missed 34 of the 126 mistakes, against 10 for Gemma 4 E2B and Buddy. For a learner a miss is the worse failure, because silence reads as approval.
The retired models were timed in August, before the instruction cache, so their times are not comparable. Buddy's 0.3 s is Gemma 4 E2B's: the same size and precision, not yet timed on its own.
Private by default · เป็นส่วนตัว
Your sentences stay on your Mac.
- On the device, offline. Apple's Translation framework and the local model both run on your Mac. Translation works with no network at all.
- The local model is opt-in. Turn it on in Settings when you want it. Without it, translation is free and still offline.
- Cloud models only if you ask. API mode is off by default. If you turn it on, your keys live in the Keychain, never in a settings file or a backup.
- No account. Nothing to sign up for, and nothing synced anywhere.
AI model
Accuracy is out of 200 sentences — 126 with a mistake in them, 74 already correct. Memory is what the app holds while running; models over 5 GB are marked. A dash means no one has measured that model yet.
API models send your text off this Mac and are billed per call. The local models never do either.
In the private beta · Buddy 1.0
Buddy, tuned for Thai finance.
Gemma 4 E2B, fine-tuned on Thai finance. It carries more of the meaning and raises fewer false alarms, at about the same speed and the same memory, and it downloads already packed.
Requirements · สิ่งที่ต้องมี
What you need.
A Mac with Apple Silicon
Built for arm64 only. The local model needs unified memory, so an Intel Mac is refused at install rather than given a feature that silently does nothing.
macOS 15 Sequoia or newer
Apple's on-device translation needs it. The first translation triggers Apple's one-time language download.
2.0 GB, once
Buddy 1.0 downloads already packed, so nothing is repacked on your Mac. It needs about 2.5 GB of free memory while loaded. Translation itself needs none of this.
Private beta · registration open · เปิดรับสมัคร
Try Languagly first.
We're looking for people to use it in their real work and tell us what works and what doesn't. It's free.
- Who: Thai speakers who write English at work — finance especially.
- Your Mac: Apple silicon (M1 or later) and macOS 15 or newer. Intel Macs can't run it.
- What we ask: use it in your daily work, and answer a few short questions in a LINE group.
A two-minute Google Form, in Thai and English. We choose testers to cover different kinds of Mac, so not everyone is picked — we'll email you if you are. Your answers are used only for this beta and deleted when it ends.