B1ade Embed
w601sxs/b1ade-embed
published May 2024 · updated Mar 2025
B1ade Embed is a 335M parameter embedding model optimized for RAG, ranking #1 in STS on the legacy MTEB leaderboard for models under 500M parameters.
specs
| Task | Feature Extraction (embedding) |
| Architecture | BERT-large-uncased based (model merge) |
| Parameters | 335M |
| License | Not specified |
about this model
b1ade-embed is a feature extraction (embedding) model optimized for retrieval-augmented generation (RAG) tasks, with 335 million parameters.
Performance Benchmarks
On the legacy MTEB leaderboard (2024), b1ade-embed ranked #1 in the Semantic Textual Similarity (STS) category among models under 500M parameters. It also placed competitively in ranking, retrieval, and classification tasks.
Architecture and Training
The model combines model merging of five base models (bert-large-uncased, WhereIsAI/UAE-Large-V1, BAAI/bge-large-en-v1.5, mixedbread-ai/mxbai-embed-large-v1, avsolatorio/GIST-large-Embedding-v0) with knowledge distillation from larger models. It uses the Safetensors format and is tagged for text-embeddings-inference.
Research Validation
A medRxiv paper evaluating 30 embedding models across clinical tasks (2.1M comparisons) highlighted b1ade-embed's strong performance in both clinical and PubMed domains, noting its high efficiency despite smaller size. On short clinical tasks (triage notes, chief complaints) it achieved a score of 27.4, competing closely with larger models.
A CEUR-WS paper demonstrated b1ade-embed's effectiveness in taxonomy enrichment for the labor market, achieving an 81% Positive Predictive Value (PPV) when combined with other models in a closed-world evaluation using ESCO's hierarchy.
Availability
Hosted on gigarouter as an OpenAI-compatible API. b1ade-embed is part of the b1ade collection of small RAG models. Monthly download count: 209.
best for
- ·Semantic Textual Similarity (STS) tasks
- ·Retrieval-Augmented Generation (RAG) pipelines
- ·Clinical and biomedical text analysis
FAQ
It is designed for RAG and excels at Semantic Textual Similarity (STS), ranking #1 in that category on the legacy MTEB leaderboard for models under 500M parameters.
It has 335 million parameters.
It accepts text input and outputs a dense vector embedding. Use the Hugging Face AutoTokenizer and AutoModel to generate embeddings.
Use the gigarouter OpenAI-compatible endpoint with your API key, specifying the model name w601sxs/b1ade-embed.
The model card does not specify a license.
We're benchmarking and onboarding B1ade Embed as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.