Hacker News front pageApr 27, 2026, 09:54 UTC+8

TurboQuant: A First-Principles Walkthrough

TurboQuant walkthrough explains compressing AI vectors to 2–4 bits per coordinate. It uses random rotation to map high-dimensional coordinates to a fixed distribution, then reuses one codebook with no scale overhead, training, or calibration.

Why it matters: HKR-H/K/R all pass: the hook is concrete, the mechanisms are new, and the cost angle is relevant. It stays below 78 because this is a technical walkthrough, not a model or product release.

Read the original ↗Export Markdown

Take

TurboQuant’s useful claim is not the headline compression ratio. It attacks the operational mess around quantization: random rotation pushes high-dimensional coordinates toward a fixed distribution, then one universal codebook handles the vectors. The page explicitly claims 2–4 bits per coordinate with no scale-factor memory, no training, and no calibration. For KV cache work, that last part matters because calibration drift across layers, heads, and models is where clean papers become messy deployments.

I don’t fully buy “without losing accuracy” from this walkthrough alone. The article gives the mechanism, browser demos, and links to TurboQuant-MSE, PolarQuant, and QJL; it does not show a hard Llama/Qwen long-context benchmark table in the provided body. Against KIVI-style KV compression, the pitch is fewer moving parts, not magic. If the inner-product bias correction fails under real attention workloads, 2-bit quickly becomes a 4-bit product story.

Source: Hacker News front page · arkaung.github.io
Published: Apr 27, 2026, 09:54 UTC+8 · Apr 27, 2026