7.7M active parameters
According to the model card, only 7,671,168 parameters run per token, while the multilingual embedding table brings the total to 105,975,168.
Open Source Model Profile · hotchpotch
bekko-embedding-v1-a8m
bekko-embedding-v1-a8m is an ultra-compact multilingual embedding model from hotchpotch. According to the model card, it uses 7.7M active parameters with CPU and browser-ready builds.
- Publisher
- hotchpotch
- Task
- sentence-similarity
- Model type
- modernbert
- License
- mit
- Library
- sentence-transformers
- Publication status
- Accepted · not indexed
Model overview
bekko-embedding-v1-a8m is published by hotchpotch as a modernbert-based sentence-similarity model. The captured configuration identifies ModernBertModel, and Safetensors metadata reports 105975168 parameters. According to the model card, it is an ultra-compact multilingual embedding model with 7.7M active parameters, fine-tuned from hotchpotch/bekko-embedding-v1-a8m-pt.
Recorded capabilities
Multilingual 8k Matryoshka embeddings
The card describes 100+ languages, 8192-token context, 384-dimension mean-pooled embeddings truncatable to 256, 128, or 64 with cosine similarity.
CPU and browser-ready builds
The card documents OpenVINO CPU inference, ONNX browser deployment with Transformers.js, and SDPA or Flash Attention 2 GPU paths.
Use cases in the source record
- Multilingual retrieval and similarity search, including truncated 256, 128, or 64-dimension Matryoshka embeddings for smaller indexes.
- Low-resource CPU and in-browser embedding experiments using the card's documented OpenVINO, ONNX, and Transformers.js paths.
Limitations and unknowns
- No independently measured evaluation results were extracted; HAKARI-Bench and MMTEB figures are publisher-reported card claims.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: hotchpotch/bekko-embedding-v1-a8m
Captured: Unknown. Processed: 2026-09-07T19:36:07.661547+00:00.
bekko-embedding-v1-a8m bekko-embedding-v1-a8m is an ultra-compact multilingual text embedding model. It has just 8M active parameters — light enough to run comfortably even on low-spec CPUs — yet its retrieval quality is comparable to models with 3–10x more active parameters. For higher retrieval quality, see the larger bekko-embedding-v1-a25m (25M active parameters). You can also try bekko right in your browser: the bekko-embedding-web demo runs the model fully client-side with Transformers.js — no server involved. For a guided overview of the models, training recipe, and results, read Bekko Embedding: how small can a multilingual…
F001F002F003F004F005F006F007F010F011F013F015F021F022F024F025F029F031F032F034F036