GTE Base EN v1.5
Alibaba-NLP/gte-base-en-v1.5
published Apr 2024 · updated Nov 2024
GTE Base EN v1.5 is an English text embedding model that supports up to 8192 tokens and achieves state-of-the-art MTEB scores in its size category.
specs
| Task | Text Embeddings |
| Architecture | Transformer++ (BERT + RoPE + GLU) |
| Parameters | 137M |
| License | Apache 2.0 |
about this model
Alibaba-NLP/gte-base-en-v1.5 is an English text embedding model that encodes input texts into dense vector representations, supporting a maximum sequence length of 8192 tokens.
The model is built on a transformer++ encoder backbone combining BERT with Rotary Position Embedding (RoPE) and Gated Linear Units (GLU). It employs attention dropout of 0 for compatibility with xformers and flash attention, and uses unpadding to eliminate computation on padding tokens, improving inference efficiency.
On the MTEB benchmark (56 tasks), gte-base-en-v1.5 achieves an average score of 64.11, competitive with larger models: bge-base-en-v1.5 (109M parameters) scores 63.55, while mxbai-embed-large-v1 (335M) scores 64.68. The model also performs well on long-context retrieval, scoring 87.44 average across five LoCo tasks, matching or exceeding larger models.
| Model | MTEB Average (56) | LoCo Average (5) |
|---|---|---|
| gte-base-en-v1.5 (137M) | 64.11 | 87.44 |
| bge-base-en-v1.5 (109M) | 63.55 | – |
| gte-large-en-v1.5 (434M) | 65.39 | 86.71 |
Licensed under Apache 2.0, this model is hosted by gigarouter as a managed API, requiring no local setup.
best for
- ·Long-context document retrieval and search
- ·Semantic similarity and clustering of English text
- ·RAG pipelines requiring up to 8192-token inputs
FAQ
The model supports up to 8192 tokens.
The output dimension is 768.
GTE Base EN v1.5 scores 64.11 on MTEB (56 tasks), outperforming bge-base-en-v1.5 which scores 63.55.
Apache 2.0.
Use the gigarouter OpenAI-compatible endpoint with your API key, sending a POST request with the input text and model name.
We're benchmarking and onboarding GTE Base EN v1.5 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.