The bistro test
Sit in a small French bistro outside Lyon at 9pm on a Tuesday. The Wi-Fi is “complimentary” but you cannot reach the captive portal page. Your roaming plan turns into a 2G crawl after 200 MB. The waiter is patient but not infinitely patient. You point at “andouillette” and ask the cloud translator on your phone what it is.
It thinks for eight seconds. It returns “andouillette.” The waiter waits. You guess. Two of those bites later, you find out exactly what andouillette is.
This article is the case for why on-device AI is not just a privacy upgrade — it is the only travel-shaped product that survives this test.
The three axes
I used to think on-device AI was a privacy story. After enough trips, I realized it is actually three stories that happen to align in the same direction.
Network reliability
Travel is a network-hostile environment. You are in basements, on trains, in captive Wi-Fi portals, in countries where your eSIM is “available” but practically a 2 kbps trickle. Cloud translation depends on stable upstream bandwidth to send the source text and stable downstream to receive the translation. Either one being flaky breaks the loop.
On-device translation removes the variable entirely. It works the same in your living room and on a Himalayan ridge.
Privacy
Cloud translation services log requests. You can read the privacy policies — most of them say “we may store and process translation requests for service improvement.” That includes your medical questions at a foreign pharmacy, your private journal entries you wanted “cleaned up,” and the awkward conversation you tried to translate with the bartender.
On-device means none of that text leaves the phone. The trust model is: the data never even reaches the trust boundary.
Latency
This is the one that people underestimate. Cloud translation round-trip on a good connection is roughly 300–500 ms. On a flaky connection it is 2–8 seconds. On-device on a Pixel 9 is sub-500 ms with no network variance.
The difference between 500 ms and 5 seconds is not “5× slower” — it is the difference between a conversation feels normal and a conversation is abandoned.
A side-by-side test
I ran the same four sentences through Google Translate (cloud) and Cove Travel (Gemma 4 E2B on-device, see Gemma model card) on a Pixel 9.
| Source | Google Translate (cloud) | Cove Travel (on-device) |
|---|---|---|
| “Could I have the bill, please?” (EN→FR) | “Puis-je avoir l’addition, s’il vous plaît?" | "L’addition, s’il vous plaît?" |
| "Is this dish spicy?” (EN→JA) | “この料理は辛いですか？" | "この料理、辛いですか？" |
| "Where is the nearest pharmacy?” (EN→ES) | “¿Dónde está la farmacia más cercana?" | "¿Dónde está la farmacia más cercana?" |
| "I have an allergy to peanuts.” (EN→ZH) | “我对花生过敏。" | "我对花生过敏。” |
The cloud and on-device versions are roughly tied. The differences are stylistic (“the bill” → “l’addition” with vs without “could I have”) rather than accuracy gaps. For travel-style sentences (under 25 words, conversational tone), both produce the same usable output.
What you do not see in the table: the cloud version took 400–800 ms on a strong connection and timed out twice on flaky ones. The on-device version took 280–410 ms every time.
When cloud still wins
I would be lying by omission if I claimed on-device wins everything. There are three cases where cloud genuinely beats on-device today:
- Long professional documents. A 10-page contract, a research paper, a technical spec — DeepL and Google Cloud’s commercial APIs handle these better than any 4B model can on phone hardware. Different category.
- Very rare languages. If you need a low-resource pair like Tagalog ↔ Wolof, neither Gemma 4 E2B nor any other on-device model is going to feel competent yet. Cloud models trained on much larger datasets still pull ahead here.
- Translation memory and team workflows. Professional translators collaborating across a team need cloud-based version control on their translations. On-device by definition does not provide that.
This is why Cove Travel is positioned as a travel companion, not as a replacement for professional translation. Use the right tool for the job.
What on-device unlocks that cloud cannot
The reverse list is shorter but more interesting. There are things on-device can do that cloud genuinely cannot:
- Working in airplane mode. This is the obvious one. Cloud literally cannot operate without a network.
- Truly private journaling. Anything you translate with a cloud service is, by definition, in someone else’s logs. On-device guarantees the trust boundary stays inside your phone.
- Predictable latency. Cloud latency is bounded by the worst path on your network. On-device latency is bounded by your CPU/NPU, which is the same in every country.
- No subscription dependency. A pay-once on-device app keeps working even if the company that built it shuts down tomorrow. A subscription cloud app stops the moment the subscription does.
That last one is underrated. The model file is on your device. Cove the company can disappear, and your installed Cove Travel still translates.
FAQ
Is on-device translation as accurate as cloud?
For travel-style sentences in high-resource language pairs, yes — within roughly 5% on standard benchmarks. For long professional documents in any language, cloud still wins.
Why does my phone feel warm when translating?
A 4B-parameter model is real compute. On a Pixel 9 with NPU acceleration, the warmth is mild and battery cost per translation is roughly the same as playing a YouTube video for 10 seconds.
What about typing accents and special characters?
That is an OS-level keyboard concern, not a translator concern. On-device translation handles whatever your keyboard produces.
Can I trust that nothing is being uploaded?
Cove Travel publishes a network audit at /en/why-offline that lists every network call the app makes. The short version: optional model updates over your Wi-Fi, optional crash reports if you opt in, and nothing else.
Companion reads: