potion-mxbai-micro

A 700KB static embedding model. Yes, really. Seven hundred kilobytes for useful sentence embeddings.

Highlights

• 68.91 avg on full MTEB English (STS + Classification + PairClassification, 25 tasks)
• 700KB total model size — fits in an email attachment
• 256 dimensions — same output dimensionality as the full model
• 80-88x faster than all-MiniLM-L6-v2 on CPU
• Pure numpy inference — no GPU needed
• Drop-in compatible with model2vec and sentence-transformers

How It Was Made

1. Start with our best 256D model (potion-mxbai-256d-v2, 70.98 avg)
2. Apply model2vec's vocabulary quantization: cluster the 29,525 token embeddings into 2,000 centroids using k-means
3. Each token maps to its nearest centroid via a token mapping table
4. The full tokenizer is preserved — all text still tokenizes correctly

The result is a 2,000-row embedding table at 256D with int8 quantization. Tokens that are semantically similar share the same embedding vector, which acts as a natural regularizer.

Benchmark Results (Full MTEB English Suite)

Evaluated on 25 tasks (10 STS, 12 Classification, 3 PairClassification), English subsets only.

Usage

from model2vec import StaticModel

model = StaticModel.from_pretrained("blobbybob/potion-mxbai-micro")
embeddings = model.encode(["Hello world", "Static embeddings are fast"])

With Sentence Transformers:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("blobbybob/potion-mxbai-micro")
embeddings = model.encode(["Hello world", "Static embeddings are fast"])

When to use this model

• You need embeddings in extremely constrained environments (embedded systems, IoT, WASM)
• You're building a browser extension or mobile app where every KB counts
• You want a fallback embedding model that loads instantly
• You need to embed millions of documents and want to minimize index storage
• Prototyping — get semantic search working in seconds, upgrade to larger models later

Model Family

Citation

@article{minishlab2024model2vec,
  author = {Tulkens, Stephan and {van Dongen}, Thomas},
  title = {Model2Vec: Fast State-of-the-Art Static Embeddings},
  year = {2024},
  url = {https://github.com/MinishLab/model2vec}
}