Qwen3 Embedding 4B

Deploy Qwen3 Embedding 4B on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. A 4B multilingual embedding model with 32K context, instruction-aware retrieval and configurable vectors from 32 to 2560 dimensions.

Text4B paramsApache 2.0Deploys in ~5 minSecure Cloud

Qwen/Qwen3-Embedding-4B

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

32K

tokens

Dimensions

2560

maximum

Parameters

dense

Running Qwen3 Embedding 4B on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Qwen3 Embedding 4B GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Qwen3 Embedding 4B vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	10 GB	1× RTX A4000	$0.20/hr
FP8	5 GB	1× RTX A4000	$0.20/hr
INT4	3 GB	1× RTX A4000	$0.20/hr

No standardized production throughput figures are published by Qwen for this model. Benchmark on the exact HexGrid GPU and serving configuration before filling these fields. Input sequence length and batching materially affect embeddings throughput.

About Qwen3 Embedding 4B

What Qwen3 Embedding 4B is built for, and where it falls short.

Qwen3-Embedding-4B is the middle-sized dense embedding model in the Qwen3 Embedding family. It uses a dual-encoder architecture and derives its semantic representation from the final EOS-token hidden state.

It supports sequences up to 32K tokens and more than 100 languages, including programming languages. Retrieval queries can include task-specific instructions to make the intended notion of relevance explicit.

The model produces vectors up to 2560 dimensions and supports Matryoshka Representation Learning, allowing deployments to request smaller output dimensions when index size or search cost matters.

What people run it for

Vector search

Create semantic vector indexes for websites, internal documentation, support content or knowledge bases.

RAG pipelines

Use as the first-stage retriever and optionally combine it with Qwen3-Reranker-4B for higher-precision candidate ordering.

Multilingual search

Represent queries and documents from multilingual or cross-language collections in a shared semantic space.

Clustering and classification

Use dense semantic representations as features for downstream grouping, similarity and classification workloads.

Strengths

Strong published multilingual and English embedding scores at half the parameter count of the 8B checkpoint
32K sequence length for long-document embedding workloads
Matryoshka output from 32 through 2560 dimensions
Instruction-aware query encoding for retrieval-specific behavior
More than 100 supported languages plus code-oriented retrieval
Supported by common embedding stacks including Sentence Transformers and vLLM

Limitations

Still significantly heavier than small embedding models designed primarily for high-throughput indexing
Using the full 2560-dimensional vector increases vector-database storage and memory relative to reduced dimensions
Qwen reports a roughly 1-5% retrieval-performance drop in many scenarios when query instructions are omitted
Smaller Matryoshka dimensions should be evaluated before production because retrieval quality can vary with dimension
Independent query/document embeddings cannot reproduce the full pairwise interaction available from a cross-encoder reranker

Quickstart

Call Qwen3-Embedding-4B through HexGrid's OpenAI-compatible embeddings API.

pythonPOST /v1/embeddings

from openai import OpenAI

client = OpenAI(
    base_url='https://<your-endpoint>/v1',
    api_key='<your-key>'
)

response = client.embeddings.create(
    model='Qwen/Qwen3-Embedding-4B',
    input='Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: Explain semantic search',
    dimensions=1024
)

embedding = response.data[0].embedding
print(len(embedding))

The HexGrid Qwen3 Embeddings endpoint supports the dimensions field. Retrieval documents should normally be embedded without the query-side instruction.

Qwen3 Embedding 4B reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

MTEB Multilingual Mean(Task)69.45
MTEB English v2 Mean(Task)74.6
C-MTEB Mean(Task)72.27
MTEB Multilingual Retrieval69.6

Qwen3 Embedding 4B specifications

Architecture and serving details for Qwen3 Embedding 4B.

PublisherQwenParameters4BLicenceApache 2.0ReleasedJune 2025

Architecture

Type: Dense dual-encoder embedding model
Backbone: Qwen3
Layers: 36
Hidden size: 2560
Attention: GQA, 32 attention heads / 8 KV heads
Pooling: Final EOS-token hidden state
Native precision: BF16

Embeddings

Context: 32K tokens
Maximum dimension: 2560
Configurable dimensions: 32-2560
MRL: Yes
Instruction aware: Yes
Languages: 100+

Serving

HexGrid endpoint: POST /v1/embeddings
Runtime: vLLM
Sentence Transformers: Supported
Transformers: >= 4.51.0
vLLM: >= 0.8.5 in Qwen examples
TEI: Supported

Qwen3 Embedding 4B frequently asked questions

Common questions about deploying Qwen3 Embedding 4B on HexGrid Cloud.

What is the default embedding size?

The model's maximum embedding dimension is 2560. Because it supports Matryoshka Representation Learning, applications can request smaller vectors down to 32 dimensions.

Should queries use an instruction?

Yes for retrieval-oriented queries. Qwen recommends a one-sentence task instruction and reports that omitting it can reduce retrieval performance by roughly 1-5% in many tested scenarios.

Do documents use the same instruction?

Qwen's retrieval examples add the instruction on the query side and embed retrieval documents as their raw text.

Can HexGrid return 1024-dimensional vectors from this model?

Yes. HexGrid documents a dimensions parameter for Qwen3 Embedding deployments. 1024 is within Qwen3-Embedding-4B's supported range.

Can I use it commercially?

Yes. Qwen released the Qwen3 Embedding series under Apache 2.0.

Deploy Qwen3 Embedding 4B today

$0.20 per hour on 1× RTX A4000, billed per minute, never shared.

Deploy Qwen3 Embedding 4B See all models