$ cat wiki/models/sl2t.md
SL2T
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-08-12 |
| Announced | 2026-08-12 |
| Context window | unknown |
| Pricing | unknown |
| License | unknown |
| Availability | Pixel 11 — Gboard and Live Transcribe (sign-to-text dictation) |
Every row but the first three is unknown, and the announcement is the reason rather | |
| than the reading. Nothing read states a parameter count, an architecture, a context | |
| window, a licence, a weights release or an API | |
(source). Pricing | |
is unknown rather than "free": the model ships inside a phone feature, and no | |
| statement about its commercial terms was read. |
Availability is a product surface, not a distribution channel. There is no
indication this model is obtainable outside the Pixel 11.
Release Date
Announced 2026-08-12 in a Google DeepMind post titled "Putting sign language AI
into users' hands"
(source). The post
arrived through the DeepMind blog feed carried in state/prefetch.json.
Ships in the same announcement — not a preview, not a waitlist. It is described as "the first of its kind to be made available within a real-world consumer product" (source).
Method
Two models, split across the device boundary (source):
- An on-device model converts camera footage into "a sort of wireframe of geometric coordinates"
- Those coordinates — not the video — are sent to Google's servers, where SL2T performs the translation
Google states the split is a privacy design. It is worth reading precisely: raw video is stated not to leave the device, and the derived pose data does. This is a hybrid system, not an on-device model, and the page records it as such rather than as "on-device sign language translation", which is how the shape of the announcement invites it to be read.
Training data: "more than 100,000 hours of multilingual sign language data", of which "approximately a quarter" is American Sign Language (source).
Benchmarks
None published. No accuracy, word-error-rate or comparative figure appears in anything read (source).
This is an accessibility product whose entire value claim is transcription quality,
announced without a single quality number. The absence is recorded rather than
filled: it appears in no sources/evals/ snapshot this repo holds, and no sign
language benchmark appears in any of them at all.
Use Cases
Stated: sign-to-text dictation in Gboard and Live Transcribe, letting deaf and hard-of-hearing users "sign to their phone anywhere where they would have normally typed" (source).
Stated audience: the "estimated 70 million people across the world who are either deaf or hard of hearing" (source).
The gap between the training set and the shipped feature is the thing to watch. The model is trained on multilingual sign data; the product translates ASL to English text only. Nothing read gives a timeline for a second language pair, and sign languages are not mutually intelligible — a multilingual training set does not make an ASL-to-English feature useful to a BSL or KSL signer.
Compared To
- SignGemma — Google previously invited public feedback on a sign language translation model under this name (source). Whether SL2T is its successor, a rename or an unrelated line is not stated in anything read, and is not inferred here
- Gemini Omni — DeepMind's general multimodal model; SL2T is a single-purpose translator shipped into a system keyboard instead
- Gemma 3n — the on-device line; SL2T's pose-extraction half is on-device but the translation half is not