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
The classifier head: fused features → 36 topic probabilities
gist.tflite
LiteRT
13MB
The same head, float32. Larger than the Core ML export because a float16 graph is not runnable: standard LiteRT/TFLite kernels cannot prepare one whose tensors are all float16, which broke the browser, Android and Linux runtimes until v2.2.0
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)
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 technologyandcreator-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.
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 (e.g. law → IAB 383News & Politics ›
Law, crafts-hobbies → IAB 248Arts and Crafts, finance → IAB 391Personal 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
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:
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). 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 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
110MB
74%
92%
bge-small-en + head
transformer embed
130MB
71%
92%
gist
static embed + n-grams + MLP
74MB
71%
91%
all-MiniLM-L6-v2 + head
transformer embed
90MB
68%
90%
potion + head
static embed
30MB
65%
89%
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. An MTEB cross-check confirms the
transformer edge is genuine static-embedding tradeoff, not a quirk of this gold.
