bekko-static-embedding-v1: GPU-free static embeddings hit retrieval SOTA, 100x faster on CPU
tomaarsen · x · 2026-10-21
hotchpotch (Yuichi Tateno) released bekko-static-embedding-v1, a multilingual static embedding model with no Transformer layers — just token lookup + mean pooling. It needs no GPU and runs 100-400x faster on CPU than multilingual-e5-small (up to 100k sentences/sec).
Key results:
- SOTA retrieval among static embeddings: 43.76 on MNanoBEIR multilingual retrieval (11 points ahead of rivals), 53.71 on English NanoBEIR
- Across MMTEB's 131 tasks it averages 50.18, within 1 point of the 12-layer Transformer multilingual-e5-small on retrieval
- Trained contrastively on 814M text pairs across 40 languages with a custom 200k-token tokenizer, batch size 16,384
More from Models
- Qwen3.8-Flash-Next runs stable at 1.66bpw exl3, fits entirely on one 5090 — Unstable_Llama · 2026-10-21
- User documents several real dangerous Gemini incidents in one thread — Regular_Manager_6235 · 2026-10-21
- DeepSeek V4 architecture reportedly redesigned: more stable and efficient, not simpler — teortaxesTex · 2026-10-21
- MiniMax's M3.1 Flash talk: a faster model isn't faster if the job isn't — AI Engineer · 2026-10-21
- OpenAI LLM flags ML paper title 'Deep Ignorance' as biological risk, refuses task — burkov · 2026-10-21
- Claude Code feels noticeably better in cloud sessions than local, users report — steffenbk · 2026-10-21