Emo: on-device emoji suggestions from text

Takes a short text string and returns the best-matching emoji. Tuned for to-dos, calendar entries, notes, and message drafts across 23 languages (including CJK, Arabic, Thai, Hindi, and more). The whole thing, model and tokenizer, is small (~5 MB on Apple via Core ML, ~11 MB via LiteRT elsewhere) and runs in well under 2 ms on device.

"Dentist appointment" → 🦷 · "réserver un vol pour Tokyo" → ✈️ · "犬の散歩" → 🐕 · "จองโรงแรม" → 🏨

Try it

All platforms ship from one repo: Desert-Ant-Labs/emo (Swift, Kotlin, and JavaScript in a single codebase).

Live demo: desert-ant-labs/emo-demo: type a phrase, get emojis, fully in your browser.
iOS / macOS / tvOS / visionOS: the Swift SDK (Swift Package Manager), with a built-in demo app. It bundles the compiled Core ML model below.
Android / Kotlin: Maven Central ai.desertant:emo with LiteRT (.tflite). The small model is bundled by default; exclude ai.desertant:emo-tflite-resources to force on-demand download or explicit-directory loading.
Node / browser (JavaScript / TypeScript): npm i @desert-ant-labs/emo @litertjs/core for browser builds, or just npm i @desert-ant-labs/emo for server-side Node. The npm package downloads the model from this repo on first use and caches it; browser inference uses LiteRT.js, and Node uses prebuilt native libraries (Core ML on macOS, LiteRT on Linux). Pass directory (Node) or modelBaseUrl (browser) to self-host / run offline.

Files

File	Format	Size	Contents
`emo.tflite`	LiteRT / TFLite (int8)	~10.2 MB	Fixed-window n-gram + masked semantic inputs, softmax `probabilities` output; runs on Android, Linux, Node, and the web (bundled by default in the Kotlin SDK; downloaded on demand by the JavaScript SDK)
`emo.mlmodelc`	Compiled Core ML	~4.6 MB	Mixed 4-/8-bit-palettized transformer, ready to load on Apple platforms (used by the Swift SDK)
`emo_tokenizer.bin`	Pruned unigram tokenizer	~0.75 MB	48k SentencePiece pieces + scores; token ids = semantic-table rows
`emo_meta.json`	JSON	tiny	emoji labels + n-gram hashing / fixed-window config the runtime needs
`emo.pt`	PyTorch checkpoint	~48 MB	Full-precision weights + semantic table + tokenizer (for retraining / other runtimes)

Older revisions (tags v0.6.0 and earlier) carry Emo.mlmodelc and emo.safetensors for SDK versions that predate the unified cross-platform migration.

Architecture

A compact two-stream classifier - no large encoder, just a tiny transformer over the semantic tokens:

Lexical stream: script-aware character/word n-grams (Latin, Han·Kana, Hangul jamo, Devanagari clusters, SE-Asian, …) hashed into a fixed multi-hash signed embedding table. Its size is independent of the number of languages.
Semantic stream: a frozen multilingual static embedding (Model2Vec potion-multilingual-128M, distilled from BAAI bge-m3), PCA-reduced to 128 dims and vocab-pruned to the 48k tokens that matter for the 22 target languages. Gives cross-lingual generalization and handles out-of-vocabulary words. The matching ~0.75 MB unigram tokenizer ships alongside (emo_tokenizer.bin).
Semantic pooling: a small 2-layer transformer encoder runs over the semantic token sequence, then an attention pool - order-aware, so it composes phrases and idioms instead of averaging tokens.
Head: a small MLP fusing the two streams into a softmax over a curated vocabulary of ~800 everyday emojis (the emojis that actually come up most across the training phrases). Trained with n-gram dropout so the head relies on the semantic stream, which is what makes it generalize across languages.

Inputs and outputs

Input: a plain text string. Best on short, intent-oriented text.
Output: a probability distribution over the ~800-emoji vocabulary; take the top-1 (or top-k). Optimized for top-1 relevance.

Languages

English, Spanish, Portuguese, French, German, Italian, Dutch, Russian, Polish, Turkish, Arabic, Chinese (Simplified & Traditional), Japanese, Korean, Hindi, Indonesian, Thai, Vietnamese, Ukrainian, Swedish, Danish, Czech.

Limitations

Tuned for short, intent-oriented text; long-form text produces noisier suggestions.
Emoji semantics are imprecise; near-ties at the top of the ranking are expected.
Per-language quality varies; lower-resource languages in the set are somewhat weaker.

Built on

minishlab/potion-multilingual-128M (MIT): semantic embedding stream (PCA-reduced, vocab-pruned derivative) + tokenizer lineage.
BAAI/bge-m3 (MIT): teacher the static embedding was distilled from.
Model2Vec (MIT): static-embedding distillation method.
Unicode CLDR emoji annotations: multilingual keyword grounding in the training data.

See THIRD_PARTY_NOTICES.md.

License

Desert Ant Labs Source-Available License. Free for most apps; a commercial license is required at scale. Full terms are at the link. Licensing: licensing@desertant.com.

Citation

@software{emo_2026,
  title  = {Emo: on-device emoji suggestions from text},
  author = {Desert Ant Labs},
  year   = {2026},
  url    = {https://huggingface.co/desert-ant-labs/emo},
}

Downloads last month: 1,040

desert-ant-labs
/

emo