# Qwen3-Embedding-0.6B (slim)

> Qwen3-Embedding-0.6B is Alibaba's open-weight text embedding model of 0.6B parameters, built on Qwen3-0.6B-Base, for retrieval, classification and clustering in over 100 languages and code, with 1,024-dimension vectors and a 32K context, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-embedding-0-6b.md (~10,400 tokens) · this version ~2,280 tokens · JSON https://www.anchorterminal.com/tools/qwen3-embedding-0-6b.json · canonical https://www.anchorterminal.com/tools/qwen3-embedding-0-6b
- 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-0-6b#pricing · reviews https://www.anchorterminal.com/tools/qwen3-embedding-0-6b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 63.5/100 · rank #23 of 100 open-weight models · #7 in Embeddings & rerankers · not agent-ready · confidence medium**


Assessment: A tiny local embedding cousin for multilingual and code retrieval, under plain Apache 2.0, that fits in under 1 GB at Q8_0 and runs in llama.cpp, Ollama, vLLM, SGLang and Transformers. Queries need a task instruction and last-token pooling. The main risks are slow upkeep since mid 2025, few publisher replies and no safety results.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Embeddings & rerankers · legal entity: Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-Embedding-0.6B`, huggingface `Qwen/Qwen3-Embedding-0.6B-GGUF`, ollama `qwen3-embedding:0.6b`
- 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-0.6B and Qwen/Qwen3-Embedding-0.6B-GGUF, read from the model cards, the files and their metadata, the Sentence Transformers and pooling configs, the tokeniser config, the QwenLM/Qwen3-Embedding GitHub repository and technical report, Qwen's blog post, Ollama's library and the runtimes' own code and documentation on 21 October 2026. Nothing was downloaded or run
- Model: A 28-layer dense Qwen3 decoder fine-tuned from Qwen/Qwen3-0.6B-Base (declared in the card's base_model), 595,776,512 parameters per the files. Text in, one vector out, taken from the last token's hidden state and L2-normalised. The config still names Qwen3ForCausalLM
- Input and output: Up to 32,768 tokens of text (the card's examples truncate at 8,192). Queries take a one-sentence task instruction as "Instruct: {task}\nQuery:{query}"; documents take none. The tokeniser appends the end-of-text token and pooling uses the last token, with left padding recommended. Output is a 1,024-dimension vector; the card says user-defined dimensions from 32 to 1,024 work (Matryoshka) but doesn't show how to ask for them
- Languages: Over 100 languages, including programming languages, per the card
- Files: One BF16 safetensors file with the tokeniser, Sentence Transformers configs and last-token pooling config, 1.21 GB for the whole repository. Qwen's GGUF repository holds Q8_0 (0.64 GB) and F16 (1.20 GB) files. No pickle files and no Python files. Hugging Face lists a SHA-256 for every LFS file. No LICENSE file in either Hugging Face repository or in the GitHub repository; the card metadata says apache-2.0
- Licence: Apache 2.0, with no use policy. The base model, Qwen3-0.6B-Base, is also Apache 2.0, so the chain adds no terms
- Runtimes: llama.cpp (the qwen3 architecture with --pooling last, documented on Qwen's GGUF card), Ollama (its own qwen3-embedding:0.6b), vLLM (listed in its pooling models at v0.32.0), SGLang (listed in its embedding models at v0.5.22, served with --is-embedding) and Transformers (qwen3). The card also documents Sentence Transformers and Text Embeddings Inference. Not in LM Studio's hub
- Recommended settings: Left padding, FlashAttention 2 with float16 where available, a task instruction on queries written in English, cosine similarity on normalised vectors, per the card
- Hardware: The card states no memory figure and points to the blog and GitHub for hardware needs, which give none. The files are 0.64 GB at Q8_0 and 1.2 GB at BF16
- Training data: The technical report describes about 150 million synthetic weakly supervised pairs generated with Qwen3-32B, then about 7 million labelled pairs (MS MARCO, NQ, HotpotQA, NLI, DuReader, T2-Ranking, SimCLUE, MIRACL, MLDR, Mr.TyDi, Multi-CPR, CodeSearchNet and others) and about 12 million synthetic pairs kept by a cosine similarity above 0.7, followed by slerp model merging. No cut-off is given

- Scores: Capability 56, Reach 90, Risk 55, Agent interface 65, Documentation and upkeep 39 · 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, Memory earns 30 because Qwen's own Q8_0 GGUF and Ollama's library build are 0.64 GB, under 1 GB with the fifth added. · Risk, File safety gets 25, safetensors and GGUF only, loaded by the runtimes' own code. · Agent interface, Read for an embedding model, the template, tool-calling, trained tool calling and structured-output lines stand for what it takes and return… · Documentation and upkeep, The card covers intended use, architecture, parameters, context and languages and misses limitations and a cut-off (14).
- Sources: 20, open questions: 5, 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-0-6b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-embedding-0-6b.svg` or a link to https://www.anchorterminal.com/tools/qwen3-embedding-0-6b 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. Start it with `ollama pull qwen3-embedding:0.6b`, then POST to /api/embed with the text in `input`
2. With llama.cpp, run `llama-server -hf Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0 --embedding --pooling last -ub 8192`; without `--pooling last` the vectors are wrong
3. Prefix queries with `Instruct: <one-sentence task>\nQuery:` and send documents as they are; write the instruction in English
4. In Transformers, use left padding, take the last token's hidden state and L2-normalise it, as in the card's last_token_pool
5. In Sentence Transformers, pass `prompt_name="query"` for queries; the prompt is stored in config_sentence_transformers.json
6. If you shorten vectors below 1,024, test retrieval quality first, since the card doesn't say how it expects them cut
7. Retrieved documents are untrusted text; embedding them doesn't remove instructions an attacker put in them

## Connect

```bash
ollama pull qwen3-embedding:0.6b
# or, with llama.cpp and Qwen's GGUF:
llama-server -hf Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0 --embedding --pooling last -ub 8192
```

```bash
curl http://localhost:11434/api/embed \
  -d '{"model": "qwen3-embedding:0.6b", "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-4B | BB | 72.9 | embed.text, embed.multilingual, embed.code, model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.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)

- ★★★☆☆ Commit messages say what changed, then three replies (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★★ 0.64 GB embeddings, Q8_0 from Qwen (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Instruction format clear, dimensions left unexplained (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Vectors out, so the exposure sits in what it indexes (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 3 to 5. They agree the files are safetensors and GGUF loaded by five runtimes' own code under Apache 2.0, and that the card documents the instruction format and pooling but not how to ask for shorter vectors. Read it as a 0.64 GB embedder with a Q8_0 build from Qwen, whose replies stopped in June 2025.