Instructions to use MisasAI/LittleLittlePotion-164kb with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Model2Vec
How to use MisasAI/LittleLittlePotion-164kb with Model2Vec:
from model2vec import StaticModel model = StaticModel.from_pretrained("MisasAI/LittleLittlePotion-164kb") embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."]) print(embeddings.shape) - Notebooks
- Google Colab
- Kaggle
LittleLittlePotion-164kb
A 164 KB embedding model. It's the even smaller sibling of LittlePotion-1BPW, if you're curious check my writeup on how this was done. squeezed with LittleBit, a compression method from Samsung Research.
This is an unofficial compressed model built on MinishLab's Model2Vec/Potion work.
How good is it?
| Size | own test queries | SciFact | STS-B | SICK-R | |
|---|---|---|---|---|---|
| original (vocab-quantized Model2Vec) | 748 KB | 0.426 | 0.373 | 0.724 | β |
| LittleLittlePotion | 164 KB | 0.454 | 0.363 | 0.716 | 0.621 |
4.5x smaller, slightly better than the original Model2vec version on the kind of queries it was trained for, and within a few percent on the public benchmarks. SciFact is search (nDCG@10); STS-B and SICK-R are sentence similarity (Spearman). And yet, it's still garbage and unusable. But at 164kb, you cannot complain. It's a proof of concept that smaller isn't inconsistent with "better". Unless of course, you count the actual base.
For comparison, LittlePotion-1BPW is 931 KB and scores higher (STS-B 0.737, SciFact 0.453).
This is a proof of concept model. Check the above linked LittlePotion-1BPW for one thaat's actually decent.
What's in here
| File | What it is |
|---|---|
littlelittlepotion-164kb.lbit |
the model (164 KB) |
tokenizer.json |
its tokenizer (4x the size of the model |
littlepotion.py |
small loader for demonstrations |
benchmark.py |
checks the numbers below yourself |
scripts/ |
everything used to make it |
How to use it
from huggingface_hub import snapshot_download
import sys
path = snapshot_download("MisasAI/LittleLittlePotion-164kb")
sys.path.insert(0, path)
from littlepotion import load
model = load(f"{path}/littlelittlepotion-164kb.lbit")
vecs = model.encode(["The cat sat on the mat.", "A kitten rested on the rug."])
print(vecs @ vecs.T) # higher = more similar
You need torch, safetensors and tokenizers. If model2vec is installed you get a normal StaticModel back, otherwise a small stand-in with the same encode().
How it was made
It started from a 748 KB Model2Vec model that had already been shrunk once: instead of every word piece having its own vector, words share 2,000 vectors, plus a lookup of which word uses which and a small per-word scale.
- The 2,000 shared vectors were compressed with LittleBit to 1 bit per weight: two small tables of +1/-1 plus a few scaling numbers.
- The word-to-vector lookup was packed into 11 bits per word (lossless).
- It was then retrained in short 250-step stages, each with a different teacher: four sizes of Ettin reranker, then bge-small. A stage was kept only if validation search improved and sentence similarity didn't drop. That stopped it overfitting.
Check it yourself
pip install -r requirements.txt
python benchmark.py
Training scripts
| Script | What it does |
|---|---|
distill_stages.py |
the staged training with reranker teachers that made this model |
micro_lb.py |
compresses the shared-vector table (simpler, similarity-only version) |
convert_generated.py |
turns a {query, positive, hard_negatives, paraphrases} jsonl into the training format |
gen_paraphrases.py |
adds filtered paraphrases to a dataset with a small local LLM |
pipeline.py |
BM25F, then dense, then Ettin reranker retrieval on SciFact, for comparison |
lbit_decode.py |
helper used by the scripts above |
The scripts expect the original 748 KB model in an orig/ folder next to them.
Credits
- Model2Vec and potion by Stephan Tulkens and Thomas van Dongen, MinishLab (MIT).
- LittleBit by Lee, Kim, You and Kim, Samsung Research, NeurIPS 2025 (arXiv:2506.13771). This is an independent implementation from the paper, not Samsung's code.
- Teachers: Ettin rerankers (
cross-encoder/ettin-reranker-*-v1, Apache-2.0) and BAAI bge-small-en-v1.5 (MIT).
@software{minishlab2024model2vec,
author = {Stephan Tulkens and {van Dongen}, Thomas},
title = {Model2Vec: Fast State-of-the-Art Static Embeddings},
year = {2024},
publisher = {Zenodo},
doi = {10.5281/zenodo.17270888},
url = {https://github.com/MinishLab/model2vec},
license = {MIT}
}
@inproceedings{lee2025littlebit,
author = {Banseok Lee and Dongkyu Kim and Youngcheon You and Youngmin Kim},
title = {LittleBit: Ultra Low-Bit Quantization via Latent Factorization},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2025},
eprint = {2506.13771},
archivePrefix = {arXiv}
}
Model tree for MisasAI/LittleLittlePotion-164kb
Base model
minishlab/potion-base-8M
from model2vec import StaticModel model = StaticModel.from_pretrained("MisasAI/LittleLittlePotion-164kb") embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."]) print(embeddings.shape)