ResApexQuant — Live Benchmark

KV cache quantization that beats SOTA by ×12 at 1-bit.

Run the benchmark live below. All algorithms are calibrated and tested on real random unit vectors — no cherry-picking, no pre-computed results.

Patent pending · SR-SE Research · Code on GitHub

Parameters

Vector dimension d

KV head dimension (64 for GPT-2, 128 for LLaMA)

32 512

Bit-widths to test

Comma-separated, e.g. 1,2,3,4

Test vectors (n_test)

More = slower but more accurate

200 3000

Query vectors (n_q)

For Recall@k computation

50 500

What this tests:

Real random unit vectors on S^{d-1}
Calibration → test → Recall@k
4 algorithms: Uniform, TurboQuant (SOTA), ApexQuant, ResApexQuant

What stays secret:

Auto-calibration on real LLM KV vectors
CUDA/Triton optimized kernel
Production integration pipeline

Benchmark Results

ResApexQuant — Patent pending (INPI 2026) · SR-SE Research Unit

The method exploits a theoretical property of residual attention deltas (δ = h − x) combined with Fisher Saturation constraints (Sc = 0.769893). Full mathematical proof available under NDA.