What newer quantization options - beyond the original scalar, product, and binary trio - has Qdrant introduced to fine-tune the compression/accuracy curve?

Qdrant has expanded its quantization menu beyond the original scalar/product/binary trio in recent releases, primarily to fill the gap between int8 (4x) and binary (32x), and to reduce the accuracy cost of binary without giving up its memory savings. The most significant additions are sub-byte quantization - schemes that use fewer than 8 bits per dimension but more than 1, such as 1.5-bit and 2-bit variants - and asymmetric quantization, where the query vector is kept in full precision while the stored vectors are quantized. Sub-byte schemes let you pick a point on the compression curve between scalar and binary, which matters because the jump from 4x to 32x is very large and most datasets do not need to go all the way. Asymmetric quantization changes the error model: by keeping the query in float32, the distance computation is asymmetric (one operand full precision, one quantized), which can substantially reduce the quantization error compared to quantizing both sides, at the cost of a slightly slower distance computation.

The mechanism for sub-byte quantization is essentially the same as scalar quantization - a linear map from float to a small integer grid - but the grid is coarser (4 or 16 levels instead of 256). The trade-off is more accuracy loss than int8 but less than binary, with memory savings in between. Asymmetric quantization is a different idea: instead of approximating both query and stored vector with the same low-precision grid, you approximate only the stored side and compute the distance using the full-precision query. This reduces the error substantially because the query is often the dominant source of error in symmetric quantization - the query is a single vector used across many comparisons, so its quantization error is systematic rather than averaging out. By keeping the query exact, you remove that systematic error. The cost is that the distance computation is no longer symmetric, so it cannot use the same optimized SIMD kernels as symmetric quantization, and it may be slower per comparison.