Gist

Platforms

iOS, macOS, tvOS, visionOS, Android, Linux, Windows, Browser, Node

Languages

101

Weights

v2.2.0

Demo

https://desertant.com/models/gist/

Install

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "4.0.0")

implementation("ai.desertant:gist:4.0.0")

npm i @desert-ant-labs/gist @litertjs/core   # browser
npm i @desert-ant-labs/gist                  # Node, prebuilt native core

Usage

import Gist

let gist = Gist()
let topics = try await gist.classify("How to start a podcast with just your iPhone")
// [Topic(slug: "technology", name: "Technology & Software", score: 0.93), ...]

let all = try await gist.scores(of: text)        // [String: Double], 36 entries

let posts = titles.map { PostTopics(topics: ..., timestampMillis: ...) }
let channel = channelTopics(posts, options: RollupOptions(topN: 5))
// [ChannelTopic(slug: "technology", share: 0.41, postCount: 12), ...]

import ai.desertant.gist.Gist

Gist(context).use { gist ->
    val topics = gist.classify("How to start a podcast with just your iPhone")
    // List<Topic>: slug, name, score
    val all = gist.scores(text)                  // Map<String, Double>
}

import { Gist } from "@desert-ant-labs/gist";           // browser
// import { Gist } from "@desert-ant-labs/gist/native"; // server-side Node

const gist = await Gist.load();
const topics = await gist.classify("How to start a podcast with just your iPhone");
// [{ slug: "technology", name: "Technology & Software", score: 0.93 }, ...]
gist.dispose();

import { Gist, channelTopics } from "@desert-ant-labs/gist";

channelTopics(posts, { topN: 5 });
// [{ slug: "technology", share: 0.41, postCount: 12 }, ...]

let gist = Gist(variant: .english)

Files

Inputs and outputs

Pass a plain text string: a title, or a title and a description. Gist works best on short text like posts, titles and descriptions.

Gist gives each of the 36 topics a probability (features [1, 8448] → topic_probs [1, 36]). Keep the top-k topics above the threshold in gist_config.json. Gist is tuned to tag each item with 2-3 topics, which you can then combine across a channel with channelTopics.

Topics and standard taxonomy

The 36 topics map to two industry-standard taxonomies, so you can roll up Gist output or join it into existing systems. One is IAB Content Taxonomy 2.2, with each node's stable integer ID. The other is Apple Podcasts categories. The full, machine-readable crosswalk ships in the Hugging Face repo as taxonomy_crosswalk.json. For example, law maps to IAB 383 News & Politics › Law, crafts-hobbies to IAB 248 Arts and Crafts, and finance to IAB 391 Personal Finance.

Five topics have no dedicated IAB 2.2 node, and the crosswalk flags them as Gist extensions. society-culture, creator-economy and outdoors-nature map to a nearest parent. history and self-improvement have no IAB node. film-tv rolls up IAB Movies and Television.

Languages

Gist tags topics in 101 languages. A 15-language spot check across Latin, Cyrillic, Arabic, CJK, Devanagari, Hebrew, Thai and Greek scripts gives 88% top-3. CJK, Arabic and Cyrillic scripts match or beat the Latin ones.

Model variants

The Hugging Face repo holds two builds of the same 36-topic model:

The English build is the same model with a smaller embedding and tokenizer. On English input, the English build gives the same topics as the multilingual build. The English build doesn't cover non-Latin scripts (CJK, Arabic, Cyrillic, …). Use the English build only when the input is reliably English or Latin script.

The JavaScript and Kotlin SDKs load only the multilingual build.

Evaluation

We measured recall on a held-out set of 572 real posts, labeled by people across the 36 topics. The LLM and the zero-shot classifiers ran zero-shot. Each embedding classifier got a light logistic head. Recall@3 matters most, because apps use the top few topics of each post.

Gist scores 91% recall@3, one point behind the best small models at 92%. Gist is 74MB, against 110-130MB for those two models, and runs in one on-device pass. Gist beats every zero-shot classifier on recall@3 by 18 points or more. Those classifiers never learned the taxonomy or the distribution of posts. Qwen2.5-7B, a cloud LLM, is the only model with a clear lead on recall@1: 79% against 71%.