What it is
Qwen3-VL-Embedding is a pair of open embedding models from Alibaba's Qwen team, released in January 2026 under Apache 2.0 in 2B and 8B sizes, with matching Qwen3-VL-Reranker models. Built on the Qwen3-VL vision-language model, they take text, images, screenshots, video, or any mix of these, and place them in one shared vector space, so a written question can find a chart, a slide or a video clip.
Both sizes read up to 32K tokens and support more than 30 languages. Vectors run up to 2,048 dimensions (2B) or 4,096 (8B) and can be cut to any size from 64 upwards. Qwen says task instructions improve most tasks by 1% to 5% and recommends writing them in English.
Alibaba Cloud Model Studio also sells a hosted qwen3-vl-embedding at $0.10 per million text tokens and $0.258 per million image or video tokens, with output sizes up to 2,560; Alibaba does not say whether it is the same checkpoint as the open 8B. The 8B recorded about 1.33 million Hugging Face downloads in the past month.
Our review
Qwen3-VL-Embedding is the open model to start with when search has to cover more than text. On MMEB-V2, a 78-dataset multimodal embedding benchmark, Qwen's own table gives the 8B 77.9 overall, ahead of ByteDance's Seed-1.6-embedding at 76.9 and every other model listed; the 2B scores 73.4. Qwen re-ran most rivals on an updated split, so treat the margins as indicative rather than settled.
The lead is not uniform. In the same table Seed-1.6-embedding beats it on video overall (67.7 to 66.1), and IFM-TTE beats it on the ViDoRe v2 page-retrieval subset (71.5 to 69.9). For collections dominated by scanned pages, it is worth testing a late-interaction retriever, which keeps many vectors per page, before settling on a single-vector model.
On text it gives up a little. Qwen reports 67.88 on multilingual MMTEB for the 8B, against 70.58 for the text-only Qwen3-Embedding-8B and 58.93 for OpenAI's text-embedding-3-large. One index for mixed content is simpler to run than separate text and image stores, and for many teams that convenience outweighs three points.
The hardware bill is modest. The 8B's 8.14 billion parameters come to roughly 16 GB in BF16, within one 24 GB card, and the 2B needs about 4 GB. Video is the expensive input, because each sampled frame becomes image tokens, and Alibaba's hosted version caps a request at five images and 20 items, which makes self-hosting the practical route for bulk video indexing.
Where it earns its place
An ML engineer at a broadcaster with a large news-footage archive
Let editors find clips by describing a scene
Embed sampled frames and transcripts with the 8B on a single 24 GB GPU and store the vectors in an existing vector database; cutting to 1,024 dimensions reduces storage by three-quarters. Video is the costly input, so GPU throughput, not licence fees, sets the budget.
A support team at a SaaS company whose help centre is full of screenshots
Answer 'where is this setting' questions from a pasted screenshot
Index articles as text and images in one space, query with the user's screenshot and rescore the top results with Qwen3-VL-Reranker-2B. Self-hosted, the only cost is GPU time; on Alibaba's API, text is $0.10 and images $0.258 per million tokens.
A head of data at a regulated bank
Adopt multimodal search under a licence legal will approve
Apache 2.0 allows commercial use with no revenue cap or usage policy, and the weights run inside the bank's network, so no document leaves it. The trade is staff time: there is no vendor SLA, and the team owns upgrades and monitoring.
Before you commit
- Setup
- pip install sentence-transformers and load Qwen/Qwen3-VL-Embedding-8B; the model card also gives Transformers (4.57 or later), vLLM and SGLang examples. Inputs get a default 'Represent the user's input.' instruction unless you pass your own.
- Integrations
- Sentence Transformers, Transformers, vLLM and SGLang per the model card. Alibaba Cloud Model Studio's DashScope API hosts a qwen3-vl-embedding with a fusion mode that merges text, images and video into one vector.
- Data and privacy
- Apache 2.0: commercial use, modification and redistribution allowed, with no revenue cap or acceptable-use policy attached to the weights. Self-hosted, no data leaves your infrastructure.
- Cost at scale
- Self-hosted, the 8B needs about 16 GB for BF16 weights, so one 24 GB GPU per replica; the 2B about 4 GB. On Alibaba's API, 10 billion text tokens cost about $1,000, with image and video tokens at $0.258 per million.
- Limits
- 32K tokens per input; instructions work best in English. Alibaba's hosted version allows 5 images and 20 items per request, 5 MB per image and 50 MB per video, with video by URL only.
What's good
- 77.9 on MMEB-V2, top of Qwen's own comparison table
- Text, images, screenshots and video in one vector space
- Apache 2.0, with matching 2B and 8B rerankers
- 8B fits on a 24 GB GPU in BF16
What's not
- Behind Seed-1.6 on video and IFM-TTE on ViDoRe v2 in Qwen's own table
- About 2.7 points below Qwen3-Embedding-8B on text
- Benchmarks largely self-reported
- Hosted API caps requests at five images
Best for
Teams self-hosting search across images, video and mixed documents under a permissive licence.
Alternatives
Qwen3-VL-Embedding: common questions
- Is Qwen3-VL-Embedding worth it?
- The open embedder to start with when one index has to hold slides, screenshots and video alongside text, with Apache 2.0 settling the licence question. We rate it 4.3 out of 5.
- How much does Qwen3-VL-Embedding cost?
- Free under Apache 2.0.
- What are the downsides of Qwen3-VL-Embedding?
- Behind Seed-1.6 on video and IFM-TTE on ViDoRe v2 in Qwen's own table; about 2.7 points below Qwen3-Embedding-8B on text; benchmarks largely self-reported; hosted API caps requests at five images.
- Who is Qwen3-VL-Embedding best for?
- Teams self-hosting search across images, video and mixed documents under a permissive licence.
- How much does Qwen3-VL-Embedding cost for a team or in production?
- Self-hosted, the 8B needs about 16 GB for BF16 weights, so one 24 GB GPU per replica; the 2B about 4 GB. On Alibaba's API, 10 billion text tokens cost about $1,000, with image and video tokens at $0.258 per million.
- Is Qwen3-VL-Embedding safe for company data?
- Apache 2.0: commercial use, modification and redistribution allowed, with no revenue cap or acceptable-use policy attached to the weights. Self-hosted, no data leaves your infrastructure.
- What are the best Qwen3-VL-Embedding alternatives?
- The closest alternatives we have reviewed are EmbeddingGemma 2, Jina Embeddings v5, Gemini Embedding 2 and Nemotron ColEmbed V2.
Answers checked Oct 22, 2026.
Make Qwen3-VL-Embedding? Show this rating on your site
