UEmbed

About

What UEmbed is

UEmbed is a decoder-only multimodal embedding model that produces both dense embeddings and SPLADE-style sparse lexical embeddings from a single causal forward pass. It supports text, image, video, and mixed-modal inputs for retrieval, multimodal search, and visual-document retrieval. News 🔥 - [2026-08-15]: UEmbed achieves state-of-the-art results on the text and agent tracks of MMEB-v3, and ranks second only to the Qwen3-VL-Embedding series among open-source models on MMEB-v2. - [2026-08-13]: We updated the inference code — UEmbed now loads natively with transformers (no trustremotecode or processor patching needed), and we added a vLLM backend for high-throughput dense and sparse embedding inference. See the runnable examples: examples/transformersexample.py and examples/vllmexample.py. Model Family Highlights - Unified dense and sparse retrieval: one checkpoint returns normalized dense vectors and sparse lexical vectors. - Multimodal inputs: text, images, videos, and mixed inputs are represented in the same retrieval space. - Sparse interpretability: sparse activations correspond to vocabulary terms and can be used with inverted indexes. - Causal-model serving compatibility: the sparse design keeps the decoder-only backbone, no conversion to a bidirectional encoder. Architecture Usage Requires a recent transformers build with Qwen3.5/Qwen3-VL support: Download the complete model repository, since sparse inference requires both sparseinfo.json and sparseweights.pt in the local model directory: Inference code is provided in the GitHub repository. Set pooling="last.normal" for dense embeddings or pooling="splade.last" for sparse embeddings. Input Format Qwen35Embedder.process accepts a list of dictionaries with the following fields: Training Data UEmbed is trained on 3.94M public samples: - E5 training data for broad text retrieval coverage. - M3 training data, using the MLDR subset. - MMEB training sets for multimodal query-document pairs. For multimodal data, hard negatives are mined with Qwen3-VL-Embedding-8B as the teacher retriever. Citation If you use UEmbed, please cite the paper: Acknowledgements Thanks to the Qwen3-VL-Embedding repo for the evaluation f

Summarised from the published model card. Read the full card on the HuggingFace links below.

Specifications

What it is

How it works

How embedding models work

Variants

Sizes & precisions

Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys uembed for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (uembed below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/embeddings \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"uembed","input":"text to embed"}'

Details

Languages, data & research

Papers

Licence

Open weights

Sources

Weights & code

Explore

More embedding models