gte-multilingual-base

gte-multilingual-base, published on 20 July 2024, is the multilingual member of Alibaba's GTE family and the one described in the company's mGTE paper, presented at EMNLP 2024. It converts text in more than 70 languages into a vector, so that a search index built from Polish documents can be queried in English and still return the right passage. The design choice the vendor argues for is the encoder. Competing retrieval models of the period were built by stripping down chat models — Alibaba's own gte-Qwen2-1.5B-instruct among them — which makes them large and slow. This one is an encoder only: 12 layers, hidden dimension 768, rotary position embeddings scaled to an 8,192-token window. The vendor claims a tenfold inference speed-up over the decoder-based approach at comparable quality. A detail worth knowing before judging its size: the model is listed at 305,369,089 parameters, but 192 million of those are the embedding table for a 250,048-token multilingual vocabulary. The transformer doing the actual work is roughly 113 million parameters — smaller than the English-only gte-modernbert-base the company shipped six months later. In multilingual models, a large parameter count often means a large alphabet rather than a large brain. Two features set it apart inside the family. The output vector can be truncated anywhere between 128 and 768 dimensions without retraining, trading index size for accuracy. And besides the usual dense vector it can produce a sparse one — a list of weighted tokens, in the manner of classical keyword search — which can be blended with the dense score at query time. No other GTE release offers both. One caution the model card itself supplies: Alibaba Cloud sells embedding and reranking services under the same brand (text-embedding-v1 through v3, gte-rerank), and the card states plainly that the models behind those paid APIs are not entirely identical to the open ones. Benchmark numbers published for one do not transfer to the other. Weights are Apache 2.0. Detailed results appear in the card only as images, with the numbers left to the paper. Adoption in the 30 days to 23 September 2026: 1,324,593 downloads and 377 likes, the second most downloaded model on Alibaba's research account.

▸Videos

No videos yet.

▸Reviews