qjl

Vector compression with TurboQuant codecs for embeddings, retrieval, and KV-cache. 10x compression, pure NumPy core — optional GPU acceleration via PyTorch (CUDA/MPS) or MLX (Metal).

Local-first agent memory in one SQLite file: sqlite-vec/QJL two-pass vector + FTS5 search, knowledge graph (optional auto capture,) Ed25519-signed entries, and MCP server. No cloud dependencies.

AI agent skill implementing Google's TurboQuant compression algorithm (ICLR 2026) — 6x KV cache memory reduction, 8x speedup, zero accuracy loss. Compatible with Claude Code, Codex CLI, and all Agent Skills-compatible tools.

Compress embeddings, retrieval vectors, and KV-cache with TurboQuant codecs for 10x smaller storage and NumPy-first AI workloads

Near-optimal vector quantization for LLM KV cache compression. Python implementation of TurboQuant (ICLR 2026) — PolarQuant + QJL for 3-bit quantization with minimal accuracy loss and up to 8x memory reduction.

Compress KV cache for LLM inference with 5.02x efficiency on NVIDIA GPUs using cuTile kernels.

C++17 CPU reference implementation of TurboQuant + QJL (Zandieh et al., ICLR 2026). Single-file, no deps. Validates paper theorems empirically.

A Rust implementation of TurboQuant - Google Research's near-lossless KV-cache compression algorithm for large language models (to be presented at ICLR 2026)