UEmbed¶

Development line: project:uembed · thread uembed
Last event: 2026-08-05 · 1 dated since 2026-08-05 · Researched: 2026-09-05 · confidence: medium

What it is¶

UEmbed is a Qwen3.5-based decoder-only embedding family for teams that would otherwise operate separate dense and sparse retrievers.

• Dense and sparse representations: produces normalized dense vectors and sparse lexical vectors in one causal forward pass.
• Multimodal encoding: encodes text, images, video, and mixed inputs for retrieval and visual-document search.
• Model family: ships as UEmbed-2B, UEmbed-4B, and UEmbed-9B; sparse inference requires the sparse_info.json and sparse_weights.pt sidecars.

Development line¶

• 2026-08-05 — UEmbed public project, code, and UEmbed-9B resources were linked. On 2026-08-05, we linked the official project page, its GitHub source repository, and the UEmbed-9B model page on Hugging Face. The dated links establish a public-facing project-and-model availability milestone, but do not establish a specific release version, benchmark, or capability claim.

What changed¶

• 2026-08-05 — UEmbed’s initial release was documented as 2B, 4B, and 9B decoder-only multimodal checkpoints that emit dense and sparse representations.
• 2026-08-13 — Inference moved to native Transformers loading without trust_remote_code or processor patching, and gained a vLLM backend for dense and sparse serving.
• 2026-08-15 — The project reported MMEB-v3 leadership on text and agent tracks and second place among open models on MMEB-v2 behind Qwen3-VL-Embedding.
• 2026-08-17 — Supervised fine-tuning support with ms-swift was added.

How to use this¶

From 2026-08-05, we can evaluate UEmbed through its official project documentation, inspect its source repository, and locate the UEmbed-9B model resource. Do not infer unverified performance or compatibility claims.

1. Install a current Qwen3.5/Qwen3-VL-compatible runtime: transformers>=5.4.0, PyTorch, qwen-vl-utils, tokenizers, Hugging Face Hub, Pillow, and NumPy. — https://github.com/Alibaba-NLP/UEmbed
2. Choose the 2B, 4B, or 9B checkpoint and download its complete repository; keep the sparse sidecar files alongside the weights. — https://github.com/Alibaba-NLP/UEmbed
3. Load Qwen35Embedder, pass dictionaries containing text, image, video, or mixed inputs, and supply a task instruction for retrieval queries when appropriate. — https://huggingface.co/Alibaba-NLP/UEmbed-9B
4. Select pooling="last.normal" for dense vectors or pooling="splade.last" for sparse lexical vectors, then send the chosen representation to the corresponding retrieval index. — https://github.com/Alibaba-NLP/UEmbed

Best practices¶

• Download the whole checkpoint repository; sparse mode will not work without both sparse sidecar files. — https://github.com/Alibaba-NLP/UEmbed
• Instantiate only the dense or sparse pooling mode required by the retrieval path instead of assuming one output format fits both indexes. — https://github.com/Alibaba-NLP/UEmbed
• Use task-specific query instructions for retrieval, and use BF16 plus FlashAttention 2 only where the deployed GPU/runtime supports it. — https://github.com/Alibaba-NLP/UEmbed

Superseded by this¶

• 2026-08-13 — Earlier inference guidance requiring trust_remote_code or processor patching is obsolete for the official native Transformers path.

Still unknown¶

• No immutable repository or model-card snapshot was found to independently prove that all three checkpoint files were downloadable on 2026-08-05; the paper says they were released, while the dated links include the 9B card.
• The 2026-08-15 benchmark claims are project-reported and were not independently reproduced.
• The Hugging Face card labels the checkpoint UEmbed-9B but displays “Model size 8B params”; the official materials do not explain the difference.

Sources¶

Agent brief¶

• Subject: project:uembed, thread uembed, 1 dated events 2026-08-05 → 2026-08-05.
• Practical note: From 2026-08-05, we can evaluate UEmbed through its official project documentation, inspect its source repository, and locate the UEmbed-9B model resource. Do not infer unverified performance or compatibility claims.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.