Instructions to use MisasAI/LittlePotion-1BPW with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Model2Vec
How to use MisasAI/LittlePotion-1BPW with Model2Vec:
from model2vec import StaticModel model = StaticModel.from_pretrained("MisasAI/LittlePotion-1BPW") embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."]) print(embeddings.shape) - Notebooks
- Google Colab
- Kaggle
LittlePotion-1BPW 931kb
potion-base-8M squeezed down to 1 bit per weight, from 15 MB to 931 KB, using LittleBit, a compression method from Samsung Research.
Want it even smaller? See LittleLittlePotion-164kb. And see how this was done.
This is an unofficial version and all credit goes to MinishLab for the original.
1BPW is impossible
Nope, JPEG manages similar compression. And actually the method is very similar to JPEG compression.
How it works: potion is basically one big table, one row of numbers per word piece. LittleBit rewrites that table as two much smaller tables made only of +1 and -1 (one bit each), plus a few scaling numbers. Squashing it like that breaks it at first, so it's retrained afterwards (on Wikipedia text, matching the similarities between sentences) until it gives nearly the same answers as the original.
How good is it?
| Size | STS-B | SICK-R | SciFact | |
|---|---|---|---|---|
| original potion-base-8M | 15 MB | 0.754 | 0.647 | 0.505 |
| LittlePotion-1BPW | 931 KB | 0.737 | 0.625 | 0.453 |
| kept | 16x smaller | 98% | 97% | 90% |
STS-B and SICK-R test "are these two sentences about the same thing?" (Spearman). SciFact tests search (nDCG@10). Search drops more because it was only retrained for similarity.
If you don't need it this small, normal 4-bit potion (about 4 MB) is basically lossless and simpler. This is just an interesting proof of concept model.
What's in here
| File | What it is |
|---|---|
littlepotion-931kb.lbit |
the model (931 KB) |
tokenizer.json |
potion's tokenizer |
littlepotion.py |
small loader |
benchmark.py |
checks the numbers below yourself |
scripts/potion_lb.py |
the script that made it |
How to use it
from huggingface_hub import snapshot_download
import sys
path = snapshot_download("MisasAI/LittlePotion-1BPW")
sys.path.insert(0, path)
from littlepotion import load
model = load(f"{path}/littlepotion-931kb.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().
Check it yourself
pip install -r requirements.txt
python benchmark.py
Make your own
scripts/potion_lb.py compresses and retrains potion-base-8M at any bit rate:
python scripts/potion_lb.py --bpw 1.0 --steps 3000
Credits
- potion-base-8M and Model2Vec by Stephan Tulkens and Thomas van Dongen, MinishLab (MIT). LittlePotion is just their work, compressed.
- 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.
@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/LittlePotion-1BPW
Base model
minishlab/potion-base-8M