Show HN: TurboQuant‑WASM — Google's vector quantization in the browser

What it is

TurboQuant‑WASM is an experimental WebAssembly build of botirk38/turboquant that brings Google's TurboQuant vector quantization into browsers and Node.js. The repo packages a WASM binary (with relaxed SIMD) and a small TypeScript API — TurboQuant.init(), encode(), decode(), dot() and dotBatch() — so you can compress, search, and score high‑dimensional vectors without leaving the client. It has a live demo showing vector search, image similarity, and even 3D Gaussian splatting compression running entirely in the browser. Neat, right? The build pipeline leans on Zig and Bun, and the wasm uses relaxed SIMD extensions such as mapping @mulAdd FMA to f32x4.relaxed_madd.

How it works and claimed performance

The implementation exposes an npm package (turboquant-wasm) and claims feature parity with the Zig reference: byte‑identical outputs for the same inputs and seed, plus golden‑value tests to verify inner‑product preservation. It has been reported that payload compression is around ~4.5 bits/dim (22‑byte header excluded) and that mean absolute error for preserved dot products is <1.0 on unit vectors at dim=128. Batch scoring is highlighted: it has been reported that a single WASM call for many vectors is about 83× faster than looping over dot() calls. Those are promising numbers, although they come from the repository’s tests and demos rather than independent benchmarks.

Why it matters (and the caveats)

Why should you care? Because heavy signal‑processing and approximate nearest neighbor tricks are moving to the edge. Want low‑latency, private vector search in the browser? This is a concrete step in that direction — and it leans on modern browser tech (Chrome 114+, Firefox 128+, Safari 18+, Node 20+ for relaxed SIMD). Caveats: building the project requires Zig 0.15.2 and Bun, and some claims are based on repository tests and the underlying Google Research TurboQuant paper (ICLR 2026), so treat published numbers as promising but preliminary. The repo is MIT‑licensed, credits botirk38’s Zig work and the TurboQuant paper, and—if you like poking at fast, low‑level ML plumbing in the browser—this one’s worth a spin.

Sources: github.com/teamchong, Hacker News