gte-multilingual-base

The gte-multilingual-base model (305M parameters) produces dense text embeddings with a default dimension of 768 and supports input lengths up to 8,192 tokens. It is designed for multilingual retrieval and representation tasks, covering over 70 languages. The model uses an encoder-only transformer architecture, offering lower inference cost and approximately 10× faster speed compared to decoder-only LLM‑based embedding models of similar capability.

Key Capabilities

• Elastic dense embeddings: output dimension can be reduced to as low as 128 while preserving downstream effectiveness, reducing storage and computation.
• Sparse vector generation: the model can also produce sparse token‑weight vectors for hybrid retrieval.
• Multilingual and cross‑lingual retrieval: evaluated on MIRACL, MLDR, MKQA, BEIR, and LoCo benchmarks.
• General text representation: results on MTEB (English, Chinese, French, Polish).

Benchmark Highlights

Achieves state-of-the-art results among models of similar size on multilingual retrieval and multi‑task representation evaluations.

Retrieval performance on MIRACL, MLDR, MKQA, BEIR, and LoCo:

Detailed results on MLDR and LoCo are available in the paper.

MTEB results across languages:

Further Details

For comprehensive experimental results and methodology, see the paper mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval.