Bekko Embedding releases two tiny multilingual retrieval models with 7.67M and 24.93M active parameters
bclavie · x · 2026-07-29
Bekko Embedding open-sources two tiny multilingual retrieval models
The article introduces bekko-embedding-v1-a8m and bekko-embedding-v1-a25m, two multilingual embedding models designed for retrieval on CPU-constrained machines.
- Active parameters: 7.67M for a8m, 24.93M for a25m
- Total parameters: 106M and 123M
- Max input length: 8,192 tokens
- MMTEB Multilingual v2 retrieval score: 56.2 for a8m, 57.5 for a25m
- Full MMTEB Multilingual v2 score: 56.7 and 58.3
- CPU throughput: 364 docs/s on a Ryzen 9 7950X for a8m, 134 docs/s for a25m
- Model size: 124 MiB and 190 MiB ONNX files
The author argues small embedding models matter because agentic search and indexing often still run locally, even when LLM inference is outsourced to frontier APIs. The post also notes that open, practical multilingual models under 30M active parameters remain rare.
More from Models
- Grok 4.5 tops a new HighWalk benchmark on Laravel commit updates — elonmusk · 2026-07-29
- GPT-5.6 Pro impresses as a code reviewer and bug hunter, says one user — dejavucoder · 2026-07-29
- From GPT-2 to KimiK3, a thread argues the story is bigger than scale — algo_diver · 2026-07-29
- Users say Anthropic’s Opus 5 has become nearly unreadable after personalization changes — himanshustwts · 2026-07-29
- Scobleizer says Grok 4.5 is the best coding model right now — Scobleizer · 2026-07-29
- Apple reportedly sues OpenAI over alleged trade-secret misuse tied to future hardware — emmanuelvivier · 2026-07-29