gte micro v4
Mihaiii/gte-micro-v4
published Apr 2024 · updated Apr 2024
A popular open embeddings model, with 162 downloads a month. gigarouter benchmarks and hosts it as an OpenAI-compatible API.
about this model
Mihaiii/gte-micro-v4 is an embedding model that produces 384-dimensional dense vector representations of English text, designed for semantic-autocomplete applications. It is a distilled version of gte-tiny (22.7 million parameters, BERT-based, mean pooling, 512-token maximum sequence length).
Performance and strengths
gte-tiny, the teacher model, offers performance comparable to gte-small at approximately half the size. gte-small achieves an MTEB average score of 61.36 across 56 tasks, with the following per-category scores:
| Category | Score |
|---|---|
| Clustering | 44.89 |
| Pair Classification | 83.54 |
| Reranking | 57.7 |
| Retrieval | 49.46 |
| STS | 82.07 |
| Summarization | 30.42 |
| Classification | 72.31 |
gte-micro-v4 is a smaller, faster distillation suited for latency-sensitive semantic-autocomplete use cases. It supports a maximum of 512 tokens per input and is optimized for English text only. A demo of the intended use case is available here.
Architecture
Model size approximately 0.07 GB (70 MB), output dimension 384, sequence length 512. Uses mean pooling over BERT contextual embeddings. Licensed under MIT.
We're benchmarking and onboarding gte micro v4 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.