README.md

---
license: other
license_name: desert-ant-labs-source-available-1.0
license_link: https://license.desertant.com/1.0
language:
- multilingual
tags:
- text
- topic-classification
- content-classification
- multi-label
- on-device
- core-ml
- tflite
- litert
- multilingual
pipeline_tag: text-classification
library_name: litert
---

<!-- card-header:start (generated from manifest.json, edit below this block) -->
# Gist

Generate topics and tags for posts and articles.

Multilingual on-device content topic tagging across a 36-topic taxonomy.

- **SDKs, install and examples:** https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/docs/models/gist.md
- **Website:** https://desertant.com/models/gist/

<!-- card-header:end -->

Drop in a title, a post or a longer description and get its topics back, from a fixed
list of **36**, across **101 languages**. A compact two-stream classifier
(static embedding + hashed n-grams), with **no transformer at inference**. The deployable model is
**74MB** (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](https://huggingface.co/spaces/desert-ant-labs/gist-demo), paste a post in any language and see its topics.

<!-- card-install:start (generated from manifest.json, edit below this block) -->
| | |
| --- | --- |
| **Platforms** | iOS, macOS, tvOS, visionOS, Android, Linux, Windows, Browser, Node |
| **Languages** | 101 |
| **Weights** | [v2.2.0](https://huggingface.co/desert-ant-labs/gist) |

## Install

**Swift** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#swift))

```swift
.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.2.0")
```

Then add the `Gist` product to your target.

**Kotlin** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#android))

```kotlin
implementation("ai.desertant:gist:3.2.0")
```

**JavaScript** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#javascript-and-typescript))

```bash
npm i @desert-ant-labs/gist @litertjs/core   # browser
npm i @desert-ant-labs/gist                  # Node, prebuilt native core
```
<!-- card-install:end -->

## Use

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

```js
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 }]
```

```swift
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 | 64MB | 101-language static embedding, the semantic feature extractor |
| `gist.mlmodelc` | Core ML | 6MB | The classifier head: fused features → 36 topic probabilities |
| `gist.tflite` | LiteRT | 13MB | The same head, float32 |
| `gist_tokenizer.bin` | Unigram | 4MB | The multilingual tokenizer |
| `gist_config.json` | JSON | tiny | Slugs, feature dims, threshold |
| `taxonomy.json` | JSON | 8KB | The 36 topics (slug, name, description, IAB + Apple category) |

## 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.

## Topics and standard taxonomy

The 36 topics map to two industry-standard taxonomies so gist output can be rolled up or joined
into existing systems: **IAB Content Taxonomy 2.2** (with each node's stable integer ID) and
**Apple Podcasts categories**. The full, machine-readable crosswalk ships in this repo as
[`taxonomy_crosswalk.json`](./taxonomy_crosswalk.json) (e.g. `law` → IAB `383` *News & Politics ›
Law*, `crafts-hobbies` → IAB `248` *Arts and Crafts*, `finance` → IAB `391` *Personal Finance*).

Five topics have no dedicated IAB 2.2 node and are flagged as gist extensions
(`society-culture`, `creator-economy`, `outdoors-nature` map to a nearest parent; `history` and
`self-improvement` have no IAB node); `film-tv` is a roll-up of IAB *Movies* + *Television*.

## Languages

Topic tagging covers **101 languages**. 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.

## Model variants

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

| Variant | Location | Size | Coverage |
|---|---|---:|---|
| **Multilingual** (default) | repo root | 74MB | 101 languages |
| **English-only** | [`en/`](./en) | **15MB** | English / Latin script only |

The English build is the same model with a smaller embedding and tokenizer, so it is **topic-identical to the multilingual model on English input**. The English build does not cover non-Latin scripts (CJK, Arabic, Cyrillic, …); use it only when the input is reliably English/Latin. The Swift SDK selects it with `Gist(variant: .english)`. The JS and Kotlin SDKs currently load the multilingual build only: variant selection has to cross the shared native ABI, which has no slot for it yet.

## 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; **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%** | n/a |
| multilingual-e5-small + head | transformer embed | 110MB | 74% | 92% |
| bge-small-en + head | transformer embed | 130MB | 71% | 92% |
| **gist** | **on-device** | **74MB** | **71%** | **91%** |
| all-MiniLM-L6-v2 + head | transformer embed | 90MB | 68% | 90% |
| mDeBERTa-v3-mnli-xnli | zero-shot NLI | 560MB | 50% | 73% |
| GLiClass-base | zero-shot | 400MB | 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.

<!-- card-footer:start (generated from manifest.json, edit above this block) -->
## License

[Desert Ant Labs Source-Available License](https://license.desertant.com/1.0). Free for most
apps, and a commercial license is required at scale. Full terms are at the link.
Licensing: <licensing@desertant.com>.

See [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md).

## Citation

```bibtex
@software{gist_2026,
  title  = {Gist: Multilingual on-device content topic tagging across a 36-topic taxonomy},
  author = {Desert Ant Labs},
  year   = {2026},
  url    = {https://huggingface.co/desert-ant-labs/gist},
}
```

---

© 2026 Desert Ant Labs · <https://desertant.com>
<!-- card-footer:end -->