Qdrant's Multilingual E5 Large ONNX Export Hits Almost 1 Million Downloads
Qdrant's ONNX port of multilingual-e5-large hits 926K downloads, giving teams a lightweight way to run a top multilingual embedder in production.
- Qdrant's ONNX port of multilingual-e5-large has crossed 900K downloads, with 400K+ last month.
- Model is a 560M-parameter XLM-R-based encoder producing 1024-dim embeddings, 512-token context.
- Scores 64.4 on English MTEB (instruct variant), beating Cohere v3 and BGE-large.
- Beats 7B Mistral-based embedders on multilingual tasks, especially low-resource languages.
- Runs on CPU via FastEmbed, no torch or GPU required.
- Weak spots: 512-token limit, code retrieval, and some native single-language benchmarks.
Qdrant’s ONNX export of multilingual-e5-large had approached one million downloads at the time of writing, including more than 400,000 during the previous month. Hugging Face counters include automated and repeated pulls, so they do not map directly to unique users, but the volume shows sustained use of the artifact.
The export packages the upstream E5 model for ONNX Runtime and Qdrant’s FastEmbed library. It preserves the upstream architecture and learned weights while replacing the PyTorch inference path with an ONNX graph. Teams can therefore run the encoder on CPUs with a smaller runtime dependency set and integrate it directly into an embedding or vector-search service.
What Qdrant actually shipped
Multilingual-e5-large is a 560 million-parameter text encoder initialized from XLM-RoBERTa-large. Its training combined weakly supervised contrastive pretraining with supervised fine-tuning, according to the technical report. The encoder maps queries and documents into vectors whose proximity represents semantic similarity across languages.
| Property | Value |
|---|---|
| Backbone | XLM-RoBERTa-large |
| Parameters | Approximately 560 million |
| Vector size | 1,024 dimensions |
| Maximum input | 512 tokens, including special tokens |
| Runtime | ONNX Runtime through FastEmbed |
| FastEmbed identifier | intfloat/multilingual-e5-large |
FastEmbed handles model resolution, tokenization, batching, pooling, and ONNX execution. The first run downloads the weights unless they are already cached, so production deployments should preload the artifact or place the cache on persistent storage.
python -m pip install fastembedfrom fastembed import TextEmbedding
model = TextEmbedding(
model_name="intfloat/multilingual-e5-large"
)
documents = [
"You should stay, study and sprint.",
"History can only prepare us to be surprised yet again.",
]
document_vectors = list(model.passage_embed(documents))
query_vectors = list(
model.query_embed("What helps us prepare for surprises?")
)This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
