gist: on-device content topic tagging

Takes a piece of text — a title, a post, or a longer description — and returns the topics it is about, from a fixed 36-topic taxonomy, across 101 languages. A compact two-stream classifier (static embedding + hashed n-grams), with no transformer at inference. The deployable model is ~74 MB (int8 vocab-pruned multilingual embedding + a small fp16 head) and runs fully on device with zero per-call cost. Multi-label by design: most items carry two or three topics, and per-item scores can be aggregated across a collection (for example into channel- or feed-level topics).

"How to film a two-person podcast with two iPhones" → technology, creator-economy · "Cómo invertir en fondos indexados" → finance · "Tips for adopting a rescue dog" → pets-animals · "投资指数基金入门" → finance

Try it

Live demo: desert-ant-labs/gist-demo — paste a post in any language and see its topics.
SDKs (Swift / Android / JavaScript): github.com/Desert-Ant-Labs/gist.

Use

Drop-in SDKs run the model on device; each pins this repo's revision.

import { Gist } from "@desert-ant-labs/gist";           // browser (wasm + LiteRT.js)
// import { Gist } from "@desert-ant-labs/gist/native"; // Node (native)
const gist = await Gist.load();
await gist.classify("How to start a podcast with just your iPhone");
// [{ slug: "technology", name: "Technology & Software", score: 0.91 },
//  { slug: "creator-economy", name: "Creator Economy & Marketing", score: 0.44 }]

import Gist
let gist = Gist()
let topics = try await gist.classify("How to start a podcast with just your iPhone")

Files

File	Format	Size	Contents
`gist_embedding.i8` + `.json`	int8 static embedding	~64 MB	101-language potion embedding (261,349 tokens × 256 dims), the semantic feature extractor
`gist.mlmodelc` / `gist.tflite`	Core ML / LiteRT	~6 MB	The classifier head: fused features → 36 topic probabilities
`gist_tokenizer.bin`	Unigram	~4 MB	The multilingual tokenizer
`gist_config.json`	JSON	tiny	Slugs, feature dims, threshold
`taxonomy.json`	JSON	~8 KB	The 36 topics (slug, name, description, IAB + Apple category)

Architecture

A compact two-stream classifier — no large encoder:

Semantic stream: a frozen multilingual static embedding (Model2Vec potion-multilingual-128M, distilled from BAAI bge-m3), pruned per-script and int8-quantized. Tokenize (Unigram), gather the token rows, mean-pool, L2-normalize. Cross-lingual by construction across 101 languages.
Lexical stream: word and character n-grams hashed into a fixed vector, capturing proper nouns and exact tokens the semantic embedding smears (names, brands, gear).
Head: a small MLP fusing the two streams ([1, 8448]) into a sigmoid over the 36-topic taxonomy, trained with class balancing so it does not default to over-represented topics.

Distilled: open instruct LLMs (Apache/MIT) label the training text; a small student learns to reproduce it. Multi-label targets teach the co-occurrences (a tutorial is technology and creator-economy). Everything except the head is pure host-side code, so the same pipeline runs identically on Apple (Core ML), Android/Linux (LiteRT), and the web (WebAssembly + LiteRT.js).

Inputs and outputs

Input: a plain text string (title, or title + description). Best on short text like posts, titles, and descriptions.
Output: a probability over the 36 topics (features [1, 8448] → topic_probs [1, 36]); take the top-k above the threshold in gist_config.json. Optimized for multi-label use — an item's 2–3 topics, optionally aggregated across a collection.

Languages

Cross-lingual by construction: the multilingual static embedding shares one representation space across 101 languages, so topic tagging transfers across all of them. A diverse 15-language spot check (across Latin, Cyrillic, Arabic, CJK, Devanagari, Hebrew, Thai, and Greek scripts) gives 88% top-3, with CJK, Arabic, and Cyrillic scripts matching or beating the Latin ones — topic classification is largely language-agnostic in the shared embedding.

Model variants

Two builds of the same 36-topic model live in this repo:

Variant	Location	Size	Coverage
Multilingual (default)	repo root	~74 MB	101 languages
English-only	`en/`	~15 MB	English / Latin script only

The English build is a vocabulary prune of the same model — a smaller int8 embedding (32,251 tokens) and tokenizer, with the same classifier head, so it is topic-identical to the multilingual model on English input (no retraining). It does not cover non-Latin scripts (CJK, Arabic, Cyrillic, …); use it only when the input is reliably English/Latin. The SDKs select it with a flag — Gist.load({ variant: "english" }) (JS), Gist(variant: .english) (Swift), Gist(context, variant = GistVariant.ENGLISH) (Kotlin).

Evaluation

Recall on a held-out set of 572 human-labeled real posts (36 topics), zero-shot for the LLMs and zero-shot classifiers. Embedding classifiers get a light logistic head trained on the same corpus; recall@3 is the product metric (downstream aggregation consumes the top few topics).

Model	Type	Size	recall@1	recall@3
Qwen2.5-7B (cloud)	LLM zero-shot	server	79%	—
multilingual-e5-small + head	transformer embed	110 MB	74%	92%
bge-small-en + head	transformer embed	130 MB	71%	92%
gist	static embed + n-grams + MLP	~74 MB	71%	91%
all-MiniLM-L6-v2 + head	transformer embed	90 MB	68%	90%
potion + head	static embed	30 MB	65%	89%
mDeBERTa-v3-mnli-xnli	zero-shot NLI	560 MB	50%	73%
GLiClass-base	zero-shot	400 MB	44%	65%

gist is tied on recall@3 with the best small models, at a fraction of the size and one on-device pass — and it beats every zero-shot classifier decisively (they never learned the taxonomy or the distribution). Only a 7B cloud LLM clearly leads on recall@1. An MTEB cross-check confirms the transformer edge is genuine static-embedding tradeoff, not a quirk of this gold.

Built on

minishlab/potion-multilingual-128M (MIT): semantic embedding stream (per-script pruned, int8) + tokenizer lineage.
BAAI/bge-m3 (MIT): teacher the static embedding was distilled from.
Model2Vec (MIT): static-embedding distillation method.

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{gist_2026,
  title  = {gist: on-device content topic tagging},
  author = {Desert Ant Labs},
  year   = {2026},
  url    = {https://huggingface.co/desert-ant-labs/gist},
}

Downloads last month: 9

desert-ant-labs
/

gist