Experimental WASM + relaxed SIMD build of botirk38/turboquant for browsers and Node.js.

Based on the paper "TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate" (Google Research, ICLR 2026).

Live Demo — vector search, image similarity, and 3D Gaussian Splatting compression running in the browser.

• npm package with embedded WASM — npm install turboquant-wasm
• Relaxed SIMD — @mulAdd FMA maps to f32x4.relaxed_madd
• SIMD-vectorized QJL sign packing/unpacking and scaling
• TypeScript API — TurboQuant.init() / encode() / decode() / dot()
• Golden-value tests — byte-identical output with the reference Zig implementation

The WASM binary uses relaxed SIMD instructions:

import { TurboQuant } from "turboquant-wasm";

const tq = await TurboQuant.init({ dim: 1024, seed: 42 });

// Compress a vector (~4.5 bits/dim, ~6x compression)
const compressed = tq.encode(myFloat32Array);

// Decode back
const decoded = tq.decode(compressed);

// Fast dot product without decoding
const score = tq.dot(queryVector, compressed);

tq.destroy();

class TurboQuant {
  static async init(config: { dim: number; seed: number }): Promise<TurboQuant>;
  encode(vector: Float32Array): Uint8Array;
  decode(compressed: Uint8Array): Float32Array;
  dot(query: Float32Array, compressed: Uint8Array): number;
  destroy(): void;
}

# Run tests
zig test -target aarch64-macos src/turboquant.zig

# Full npm build (zig -> wasm-opt -> base64 embed -> bun + tsc)
bun run build

# Build WASM only
bun run build:zig

Requires Zig 0.15.2 and Bun.

Encoding preserves inner products — verified by golden-value tests and distortion bounds:

• MSE decreases with dimension (unit vectors)
• Bits/dim is ~4.5 (payload only, excluding 22-byte header)
• Dot product preservation — mean absolute error < 1.0 for unit vectors at dim=128
• Bit-identical output with botirk38/turboquant for same input + seed

MIT