**Binary quantization** stores each vector as a string of bits: one bit per dimension, set to 1 if that dimension is positive and 0 otherwise. A 1024-dim float32 vector (4096 bytes) becomes 128 bytes, a 32x compression.
The payoff is speed. Distance between two binary vectors is the **Hamming distance**: the number of bits that differ. Modern CPUs compute this as one `popcount(a XOR b)` instruction per 64-bit chunk, dramatically faster than float multiplies.
Product Quantization is a more sophisticated competitor. PQ splits each vector into chunks, learns a 256-entry codebook per chunk position, and stores one byte per chunk. PQ at 32 bytes per vector compresses 1024-dim float32 by 128x: more aggressive than binary quantization but slower per distance comparison.
The modern winner depends on the embedding model. If the embedder was trained to be quantization friendly (Matryoshka representation learning, BGE-M3, voyage-binary), binary quantization holds recall at or above 95 percent of full precision. If the embedder is a generic model that does not survive the bit truncation, PQ is the safer pick.
**Binary quantization** wins when the embedding model was specifically designed to survive it. The catch is the model, not the database side.
Mechanically, BQ is one bit per dimension: `1` if the float is positive, `0` if negative. A 1024-dim vector becomes 128 bytes. Distance is Hamming distance, computed via `popcount(a XOR b)`: a single CPU instruction over 64-bit chunks. That makes BQ both the most aggressive compression and the fastest distance computation of any production grade technique.
The historical problem was recall. With ordinary embeddings, throwing away everything except the sign destroyed retrieval quality. PQ, which learns a 256-codeword codebook per sub-vector, preserved more information at the cost of being slower (table lookups per distance) and requiring an offline codebook training step.
**What changed in 2024-2026 is the embedding side.** Matryoshka representation learning produces embeddings whose prefixes remain informative, and models like BGE-M3 and voyage-binary include binary aware fine-tuning that explicitly preserves sign only signal. With these models, binary quantization holds recall within 1-2 percent of full precision.
The decision rule: quantization friendly embedding + cost sensitive RAG goes to BQ. Billion scale corpus with full precision recall demands goes to IVF-PQ. The other options in the question are distractors (corpus size, embedding dimension, blanket claims that PQ is dead) that conflate unrelated knobs.
Quick reference
TL;DR
Binary quantization wins when the embedding model was trained to be quantization friendly (Matryoshka, binary aware). Distance becomes one popcount per 64-bit chunk.
Imagine you have a million pictures and you want to store each one as a tiny barcode so you can quickly compare them. **The chunked codebook method** (called PQ) is like splitting each picture into pieces, looking up each piece in a custom 256-entry catalogue, and writing down 32 small catalogue numbers. **The binary method** (called BQ) is even simpler: each pixel becomes either a 1 or a 0 based on whether it is positive or negative. Comparing two barcodes is then just counting how many bits differ, a single CPU instruction handles 64 bits at once. The catch used to be that throwing away that much information ruined search quality. But modern meaning vector models are now specifically trained so the signs of their numbers alone carry enough signal, which is why the binary method went from niche to default in 2024-2026.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Embedding model is not quantization aware and you cannot swap it Bit truncation destroys recall on vanilla embeddings; the recall hit is 20+ points without a Matryoshka or binary trained model.
Hard recall requirement above 0.99 with no rescoring budget BQ even with quantization friendly embeddings caps around 0.96-0.98; clearing 0.99 needs full precision re-rank or exact search on a filtered subset.
Corpus under 1M vectors The storage win of BQ over float matters at hundreds of millions of vectors; at 1M, float32 fits in 4GB and the compression complexity is not worth it.
Quick scan · 2–4 min
Everything important, quickly.
5 min: BQ mechanics, Hamming via popcount, why old embeddings broke it, what Matryoshka and binary aware training fixed, and the two stage retrieval production pattern.
Side-by-side
Property
Binary quantization
Product quantization
Bytes per 1024-dim vector
128 (32x compression)
32-64 (64-128x compression)
Distance op
popcount(XOR), 1 CPU inst per 64 bits
Table lookups per sub-vector
Build time cost
Zero (just take signs)
k-means codebook training
Recall ceiling (with ideal model)
~95-98% of float
~92-96% of float
Best embedding models
Matryoshka, BGE-M3, voyage-binary
Any modern model
Sweet spot scale
10M to 1B with rescoring
100M to 10B
Real-world examples
Real products, models, and research that use this idea.
1Qdrant Cloud ships binary quantization as a per collection option with optional rescoring against full vectors.
2Weaviate added binary quantization in 2024 and pairs it with HNSW for cost first deployments.
3Vespa supports BQ for large scale ranking workloads and exposes Hamming distance natively.
4BGE-M3 from BAAI ships with explicit binary aware fine-tuning and is the most cited open quantization friendly embedder.
5voyage-binary from Voyage AI is a commercial embedding model designed end to end for binary quantization workflows.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does the two stage retrieval pattern recover precision after a BQ first pass?
Sign in or go Pro to see the approach for this follow-up.
QWhat is Matryoshka representation learning and why does it help BQ specifically?
Sign in or go Pro to see the approach for this follow-up.
QCan you combine BQ with IVF the way PQ composes with IVF?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating binary quantization as a universal swap in for PQ. Without a quantization aware embedding model the recall collapses; it is the model side that flipped the decision.