Qwen3 embeddings and rerankers: 0.6B vs 4B, text vs VL, and where ColQwen fits
Qwen3 is the first open embedding family where the long-context claim is real: 32,768 tokens of input, embedding and reranker alike, Apache 2.0, released June 2025 (paper: arXiv 2506.05176). The 8B embedding model entered the MTEB multilingual leaderboard at number one. The family has since grown vision-language variants (January 2026). This guide covers what each member is for and what the sizes actually cost to run.
The family
| Model | Params | Dims | Context | Cost $/1M tok | Latency p50 |
|---|---|---|---|---|---|
| Qwen3-Embedding-0.6B | 596M | 1,024 | 32k | $0.011 | 157 ms |
| Qwen3-Embedding-4B | 4.0B | 2,560 | 32k | $0.039 | 464 ms |
| Qwen3-Reranker-0.6B | 596M | n/a | 32k | $0.151 | 65 ms |
| Qwen3-Reranker-4B | 4.0B | n/a | 32k | $0.050 | 580 ms |
| Qwen3-VL-Embedding-2B | 2.1B | 2,048 | 32k | $0.450 | see catalog |
| Qwen3-VL-Reranker-2B | 2.1B | n/a | 32k | $0.319 | 35 ms |
Serving numbers are from our catalog, measured on an L4. An 8B tier exists for both text families (4,096 dims); we list catalog-served models here.
Three properties define the family beyond context length. Dimensions are Matryoshka-style: user-configurable output size (32 to 1,024 on the 0.6B, more headroom on the larger tiers), so you can trade index size against quality without switching models. Queries are instruction-aware: the card is explicit that skipping the query-side instruction costs roughly 1 to 5% retrieval quality, and recommends writing instructions in English even for non-English corpora. And the rerankers score by reading the yes/no token logits of a generative model, which is why their integration differs from classic cross-encoders (left padding, and the card recommends flash attention 2).
0.6B or 4B?
Start with the 0.6B. That is the family’s honest value tier, and on our benchmarks it is a serious model, ahead of e5-large-v2 on most retrieval rows at a comparable price:
| Benchmark (NDCG@10) | Qwen3-Embedding-0.6B | e5-large-v2 | stella_en_400M_v5 |
|---|---|---|---|
| FiQA2018 | 0.473 | 0.351 | 0.559 |
| NFCorpus | 0.369 | 0.372 | 0.413 |
| SciFact | 0.695 | 0.678 | 0.784 |
| StackOverflowQA | 0.890 | 0.899 | 0.867 |
| SCIDOCS | 0.244 | 0.140 | 0.250 |
Stella still wins several English rows at a third of the price ($0.0082 vs $0.011 per million tokens); the reasons to pick Qwen3-0.6B over it are long inputs, multilingual coverage, and the adjustable dimensions. When your documents fit in 512 tokens and are English, stella remains our recommendation.
The 4B embedding model quadruples serving cost and roughly triples latency. Our published 4B benchmark coverage is thinner (three rows; on NanoFiQA2018 it scores 0.688 NDCG@10 against the 0.6B’s 0.654), so we will not sell you a big quality claim we have not measured broadly. Benchmark it on your own corpus before paying the 4x; the evals docs show how to run that comparison on SIE.
On the reranker side the same logic inverts. The 4B reranker outscores the 0.6B on AskUbuntuDupQuestions (0.695 vs 0.654 NDCG@10) and, at $0.050 per million tokens as served, is the cheaper of the two per token in our catalog; reranker cost depends on candidate volume rather than corpus size, so the economics differ from embeddings. If you rerank 50 candidates per query rather than thousands, the 4B is affordable and better.
The VL variants and ColQwen
Qwen3-VL-Embedding-2B embeds text, images, screenshots, and video frames into one 2,048-dim space. On Flickr30k image-to-text retrieval it scores 0.875 NDCG@10, between CLIP ViT-H (0.862) and SigLIP so400m (0.900), with 32k text context neither CLIP nor SigLIP can touch. For caption-style image search, SigLIP scores better and has higher throughput ($0.493 vs $0.450 per million tokens is close, with Qwen3-VL slightly cheaper per token); the VL model earns its keep when text and visual content mix in one corpus, or when inputs exceed a caption.
ColQwen2.5 (vidore/colqwen2.5-v0.2, from the ColPali line of work, arXiv 2407.01449) is a different tool: a late-interaction visual document retriever. It embeds page images as grids of 128-dim multi-vectors and matches query tokens against page regions, which makes it strong at “find the page with this table” retrieval where OCR-then-embed pipelines lose the layout. Two cautions from the primary sources: it is built on Qwen2.5-VL-3B-Instruct, whose weights carry the Qwen Research License, with only the trained adapters under MIT; check that against your commercial requirements. And multi-vector output needs a vector database that supports MaxSim scoring; our Qdrant and Weaviate integrations both handle multi-vector, documented in the integrations docs.
Serving notes
Everything above runs behind SIE’s /encode and /score endpoints. Two family-specific notes. Budget for the instruction tokens: query-side instructions add length to every request, which is billed and adds latency, and this is by design, since the same property buys the 1 to 5% quality margin. And use the Matryoshka dimensions deliberately; halving stored dimensions halves your index memory for a measurable, usually small, quality cost. Measure it on your data with the evals tooling rather than guessing.