Show HN: TurboQuant-WASM – Google's vector quantization in the browser
google vector-db
| Source: HN | Original article
Google Research has open‑sourced a WebAssembly (WASM) version of its TurboQuant vector‑quantization algorithm, letting developers run the compression and dot‑product primitives directly in the browser or in Node.js. The new repo, teamchong/turboquant‑wasm, ships a SIMD‑enabled implementation that packs embeddings to three bits per dimension, achieving roughly six‑fold size reductions while preserving dot‑product fidelity. It requires “relaxed SIMD” support – Chrome 114+, Firefox 128+, Safari 18+ and Node 20+ – and exposes just three functions: encode(), decode() and dot().
TurboQuant first entered the spotlight at ICLR 2026, where Google presented it as a near‑optimal online quantizer for LLM key‑value (KV) cache compression and vector search. In our April 4 coverage we noted its promise for breaking the AI memory wall; the WASM port now translates that promise into a practical tool for client‑side AI workloads. By shrinking embedding tables from 7.3 MB to about 1.2 MB and allowing searches on the compressed data without decompression, the library cuts bandwidth, reduces memory pressure, and speeds up inference on edge devices.
The move matters because it lowers the barrier for web‑based AI services that rely on large vector stores, such as semantic search, recommendation engines and on‑device LLM assistants. Developers can embed the compressor in single‑page apps, keep user data local for privacy, and avoid costly round‑trips to cloud back‑ends. The approach also dovetails with broader industry efforts to make AI models more efficient, a theme echoed in recent discussions about Google’s TurboQuant compression and the ongoing quest to demolish the AI memory wall.
What to watch next: Google may integrate TurboQuant into TensorFlow.js or Chrome’s upcoming AI runtime, and other open‑source projects are already building PyTorch and Rust bindings. Benchmarks comparing browser‑based compression against server‑side pipelines will reveal real‑world performance gains, while standards bodies could consider exposing quantization as a native Web API. Keep an eye on how quickly the ecosystem adopts this tool and whether it reshapes the economics of web‑scale vector search.
Sources
Back to AIPULSEN