# Qwen3-Embedding-4B (slim)

> Qwen3-Embedding-4B is Alibaba's open-weight text embedding model built on Qwen3-4B-Base, returning vectors of up to 2,560 dimensions for retrieval, code search and clustering in over 100 languages, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-embedding-4b.md (~10,550 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/qwen3-embedding-4b.json · canonical https://www.anchorterminal.com/tools/qwen3-embedding-4b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/qwen3-embedding-4b#pricing · reviews https://www.anchorterminal.com/tools/qwen3-embedding-4b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**BB · 72.9/100 · rank #3 of 100 open-weight models · #3 in Embeddings & rerankers · agent-ready · confidence medium**


Assessment: A small multilingual text embedding cousin for retrieval and code search that runs in about 3 GB from Alibaba's own 4-bit GGUF, loads in every major runtime under Apache 2.0 and has two hosts, DeepInfra and SiliconFlow. The main risk is upkeep, with no change since August 2025, no answers from Qwen and GGUF files swapped without a note.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Embeddings & rerankers · legal entity: Alibaba Cloud (named in the copyright line of the 8B sibling's Apache 2.0 licence; this repository has no LICENSE file) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-Embedding-4B`, huggingface `Qwen/Qwen3-Embedding-4B-GGUF`, ollama `qwen3-embedding:4b`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Qwen/Qwen3-Embedding-4B and Qwen/Qwen3-Embedding-4B-GGUF, read from the model cards, the files and their metadata, the configs, Qwen's GitHub repository and technical report and the runtimes' own code and docs on 21 October 2026. Nothing was downloaded or run. No hosted Qwen API for this model is listed here
- Model: A 36-layer dense decoder trained from Qwen3-4B-Base, with 4.02B parameters by Hugging Face's count (4B on the card) and tied input and output embeddings. Text in, one vector out, taken from the last token's hidden state and L2-normalised. Up to 2,560 dimensions, and Matryoshka training lets you keep the first 32 to 2,560
- Input and output: Queries take a one-sentence task instruction in the form Instruct then the task, a newline, then Query and the text. Documents go in with no instruction. The tokenizer appends the end-of-text token itself since 6 June 2025, padding goes on the left, and the card says leaving the instruction off typically costs 1 to 5 per cent in retrieval. Sentence Transformers ships the web-search instruction as the query prompt and cosine as the similarity
- Languages: Over 100 languages per the card. Qwen's GitHub README lists them by family, plus programming languages for code retrieval. The card advises writing instructions in English
- Files: Two safetensors shards in BF16 (8.1 GB with the tokenizer files) and a separate GGUF repository with Q4_K_M, Q5_0, Q5_K_M, Q6_K, Q8_0 and F16 files. No pickle files and no Python files since modeling.py was deleted on 5 June 2025. Hugging Face lists a SHA-256 for every shard and GGUF file. No LICENSE file
- Licence: Apache 2.0 per the card's metadata and Qwen's blog, with no LICENSE file in the repository or in QwenLM/Qwen3-Embedding. No acceptable use policy. The base, Qwen3-4B-Base, is also Apache 2.0
- Runtimes: Released support in Transformers (qwen3, from 4.51.0), Sentence Transformers, vLLM (Qwen3ForCausalLM in its embedding table), SGLang (its embedding docs use this model), llama.cpp (the qwen3 architecture, with --pooling last) and Ollama's library. The card also gives Text Embeddings Inference 1.7.2 commands for GPU and CPU
- Context: 32K tokens per the card and the technical report. config.json keeps the base model's 40,960 positions, which Ollama's library shows as 40K, and an open pull request from 5 June 2025 flags the mismatch. The card's Transformers example truncates at 8,192
- Hardware: The card points to the blog for hardware requirements, and the blog gives none. The BF16 files are 8.1 GB and Qwen's Q4_K_M GGUF is 2.5 GB
- Training data: The technical report describes about 150M synthetic pairs generated with Qwen3-32B from Qwen3's pretraining corpus, then about 7M labelled pairs (MS MARCO, NQ, HotpotQA, NLI, DuReader, MIRACL, CodeSearchNet and others) and about 12M synthetic pairs kept at a cosine similarity above 0.7, then model merging. No knowledge cut-off is given

- Scores: Capability 81, Reach 100, Risk 55, Agent interface 65, Documentation and upkeep 21 · total over the 7 assessed categories
- Why: Capability, From 4 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Full marks. · Risk, Safetensors and GGUF only, loaded with no code from the repository (25). · Agent interface, Read for an embedding model, the tool-calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card misses limitations, named languages and a cut-off (11).
- Sources: 24, open questions: 6, both in the full twin
- Capabilities: embed.text, embed.multilingual, embed.code, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-embedding-4b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-embedding-4b.svg` or a link to https://www.anchorterminal.com/tools/qwen3-embedding-4b from a page on qwen.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start is `ollama pull qwen3-embedding:4b`, about 2.5 GB, then POST to /api/embed
2. With llama.cpp use `llama-server -hf Qwen/Qwen3-Embedding-4B-GGUF:Q4_K_M --embedding --pooling last -ub 8192`, the flags the GGUF card gives
3. Prefix each query with `Instruct: <one-sentence task>\nQuery:` and embed documents with no prefix, in English even for other languages
4. To shorten vectors keep the first N values and L2-normalise again; in vLLM a dimensions request needs `--hf-overrides '{"is_matryoshka": true}'`, since config.json doesn't declare it
5. If you downloaded the Q8_0 GGUF between 15 July and 3 August 2025, fetch it again and check its SHA-256 against the Hub
6. Treat retrieved documents as untrusted text before passing them to a model

## Connect

```bash
ollama pull qwen3-embedding:4b
# or with Qwen's own GGUF and llama.cpp, per the GGUF card:
llama-server -hf Qwen/Qwen3-Embedding-4B-GGUF:Q4_K_M --embedding --pooling last -ub 8192
# or with SGLang, as its embedding docs show:
sglang serve --model-path Qwen/Qwen3-Embedding-4B --is-embedding
```

```bash
curl http://localhost:11434/api/embed \
  -d '{"model": "qwen3-embedding:4b", "input": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:What is the capital of China?"}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | embed.text, embed.multilingual, embed.code, model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Qwen3-Embedding-0.6B | B | 63.5 | embed.text, embed.multilingual, embed.code, model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-0-6b.min.md |
| EmbeddingGemma 300M | D | 50.7 | embed.text, embed.multilingual, embed.code, model.open-weights | https://www.anchorterminal.com/tools/embeddinggemma-300m.min.md |
| Gemini Embedding | BB | 70.6 | embed.text, embed.code, embed.multilingual | https://www.anchorterminal.com/tools/gemini-embedding.min.md |
| Qwen3-VL-Embedding-8B | B | 62.1 | embed.text, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b.min.md |


## Panel reviews (4, average 3.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ A Q8_0 GGUF briefly a third of its size (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★★ $0.02 per million tokens hosted, 2.5 GB local (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ Shorter vectors documented in the reference code (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ modeling.py gone, trust_remote_code still in the examples (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 5. They agree the repository has held no code since June 2025 and that Qwen documents the instruction format, pooling and shorter vectors, while nothing that runs the model has changed in 444 days and no question has a Qwen reply. Read it as a strong, cheap embedder at 2.5 GB or $0.02 per million tokens, with GGUF files once replaced in place.