Open-source Luxical embeddings classify 5k docs/sec on a laptop CPU at ~1/100th LLM cost
josh_wills · x · 2026-10-29
DatologyAI open-sourced Luxical (Apache-2.0), fast lexical embeddings built on token counts, TF-IDF and compact sparse-to-dense MLPs.
The accompanying discussion pitches a cheap alternative to LLM-based document classification: an open-source embedding model plus a sklearn-style classifier on top is almost always cheaper and more accurate. Benchmarks in the image show Luxical embed + MLP hitting 6,787 docs/s (23.1 MiB/s) on an M4 Max CPU — about 14x the throughput of the FineWebEdu scorer on an A10G GPU (478 docs/s). Luke Merrick adds that a two-layer ReLU net on Luxical embeddings classifies 5k docs/sec on laptop CPUs using OSS only, at 100x lower cost.
Note: the repo is no longer actively maintained and is provided for reproducibility and research.
More from coding & agent
- Multi-agent coding setup shown: Replit Agent orchestrating Codex and Claude Code in parallel — amasad · 2026-10-30
- Promptable Products: physical devices you reprogram by talking to them, launching Nov 2026 — genmon · 2026-10-30
- Building the system that displaces GitHub: agent-native CI/CD after three iterations since 2024 — kieranklaassen · 2026-10-30
- Indie dev ships an MCP server exposing 40 US government per diem and allowance tools — allowancesapiteam · 2026-10-30
- Deepgram acquires Daily, maker of open-source voice agent framework Pipecat — schwentker · 2026-10-30
- Building your own MCPs: developers share why they backed off — AaGho_ · 2026-10-30