How Weaviate 1.39 Trades Vector Memory for Recall
A new engineering explainer shows how 4-bit rotational quantization reduces vector memory, when centering helps, and why rescoring still matters.
Edited by Tyronne Panaino
Weaviate published a September 17, 2026 engineering explanation of the 4-bit rotational quantization option in version 1.39. The feature compresses vector representations more aggressively than the product's 8-bit rotational quantizer, with the goal of lowering memory use while preserving enough recall for approximate nearest-neighbour search.
This is a learning piece about an already shipped 1.39 capability, not a claim that Weaviate released a new version on September 17. The useful change for operators is a `bits: 4` setting on the existing rotational quantizer. Weaviate Cloud continues to default to 8-bit, so teams must opt into the tighter representation and evaluate the result for their own data.
Why rotation comes before compression
Quantization stores a compact approximation of each high-dimensional vector. Weaviate's approach first applies a random rotation intended to spread values more evenly across the available interval. The 1.39 implementation uses Fast Walsh-Hadamard Transforms and SIMD instructions to make that rotation and later distance calculations faster on supported processors.
Four-bit codes use half the bits of an 8-bit code, but total database memory does not fall by half because the search graph and its metadata still consume space. In the vendor's one-million-vector test, the move to the 4-bit option reduced the measured heap by about 45% compared with the earlier 8-bit configuration. That result is specific to Weaviate's dataset, index and measurement setup.
Centering changes the recall tradeoff
Weaviate also offers an optional centered mode. It estimates a mean vector from a sample, subtracts that mean before encoding and adjusts queries in the same way. The company reports that centering improved recall on several of its test datasets, particularly where embedding values were unevenly distributed.
The important caveat is that centering is not universally helpful. Some embedding models are already regularized around a mean, so Weaviate leaves the feature opt-in. A team should compare centered and uncentered configurations against an exact-search ground truth drawn from its own corpus instead of assuming the vendor's ranking transfers unchanged.
Rescoring remains another control. A search can retrieve a wider set of candidates from compressed codes and then recompute the leading candidates with uncompressed vectors. Weaviate's tests show this can recover much of the accuracy lost to tighter compression, but retaining original vectors for rescoring changes the storage and memory calculation.
What the scale tests establish
The company's brute-force experiment reported broadly stable quantizer recall as its sampled corpus grew from one million to 250 million vectors. That finding concerns the quantizer itself. Weaviate explicitly notes that approximate-nearest-neighbour index parameters can still degrade with scale and recommends appropriate sharding.
Its benchmark tables also compare encoding speed, distance kernels, import time and recall across hardware and datasets. They are useful engineering evidence, but they were produced by the vendor and were not independently reproduced in the material reviewed here. Workload shape, vector dimensions, processor instructions, search parameters and target recall can all change the real result.
Who should test it
The 4-bit option is most relevant to teams whose vector codes are a meaningful part of their memory bill and that can tolerate a measured recall tradeoff. It is less compelling when graph metadata dominates, when the dataset is small, or when an application requires the highest possible unrescored recall.
A sensible evaluation keeps the query set and ground truth fixed, measures memory and latency at the same recall target, and checks import cost as well as steady-state search. That turns the 4-bit setting from a headline saving into an application-specific decision.
Status
Learning. The feature and benchmark methodology are documented by Weaviate. Internal confidence is medium because the performance and memory results are vendor measurements rather than independent tests.
Sources
Update note: Last reviewed 2026-09-22. We will revise this explainer if Weaviate changes the 1.39 configuration guidance or publishes materially different measurements.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.