Qwen3 Embedding 8B

Deploy Qwen3 Embedding 8B on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. Qwen's 8B multilingual text embedding model, combining 32K context with instruction-aware retrieval and configurable vectors from 32 to 4096 dimensions.

Text8B paramsApache 2.0Deploys in ~5 minSecure Cloud

Qwen/Qwen3-Embedding-8B

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

32K

tokens

Dimensions

4096

maximum

Parameters

dense

Running Qwen3 Embedding 8B 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 8B GPU sizing and cost

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

Qwen3 Embedding 8B vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	20 GB	1× RTX 4090	$0.45/hr
FP8	10 GB	1× RTX A4000	$0.20/hr
INT4	5 GB	1× RTX A4000	$0.20/hr

Qwen does not publish standardized throughput figures suitable for this field. Native model weights are BF16. Embedding throughput depends strongly on input length, batch size, requested vector dimension, GPU and serving engine.

About Qwen3 Embedding 8B

What Qwen3 Embedding 8B is built for, and where it falls short.

Qwen3-Embedding-8B is the largest text embedding checkpoint in Qwen's original Qwen3 Embedding family. It uses a dual-encoder design: each query or document is processed independently and represented by the hidden state corresponding to the final EOS token.

The model supports 32K-token sequences, more than 100 languages and programming languages, and instruction-aware query embeddings for retrieval. Its default maximum vector size is 4096 dimensions.

Matryoshka Representation Learning allows applications to request smaller embeddings down to 32 dimensions, letting deployments trade vector storage and search cost against retrieval quality without switching checkpoints.

What people run it for

Semantic search

Build dense vector indexes for natural-language search over documents, knowledge bases and product or content catalogs.

RAG retrieval

Embed application queries and document chunks to provide first-stage retrieval for retrieval-augmented generation.

Multilingual retrieval

Search collections containing many languages or retrieve relevant material across language boundaries.

Code retrieval

Embed natural-language queries and code-oriented content for semantic code search and repository retrieval.

Strengths

70.58 Mean(Task) on Qwen's published multilingual MTEB evaluation
Strong multilingual, cross-lingual and code-retrieval coverage
Up to 4096-dimensional embeddings with Matryoshka dimension reduction
32K context supports long documents and document chunks
Queries can include task instructions while retrieval documents can be embedded directly
Compatible with Sentence Transformers, Transformers, vLLM and Text Embeddings Inference

Limitations

8B parameters make it substantially heavier than conventional sub-billion-parameter embedding models
Maximum-dimension 4096-element vectors require more index memory and bandwidth than smaller embedding sizes
Qwen reports that omitting retrieval instructions can reduce performance by roughly 1-5% in many retrieval scenarios
Reducing the vector dimension can change retrieval quality and should be evaluated on the target corpus
An embedding model performs independent encoding and therefore cannot model full query-document cross-attention the way a reranker can

Quickstart

Deploy through HexGrid and use the OpenAI-compatible embeddings endpoint. Retrieval queries should include a task instruction; documents normally do not.

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-8B',
    input='Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: How does retrieval augmented generation work?',
    dimensions=1024
)

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

HexGrid documents Qwen3 Embedding through /v1/embeddings and exposes the dimensions parameter for Matryoshka-capable checkpoints.

Qwen3 Embedding 8B reported benchmarks

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

MTEB Multilingual Mean(Task)70.58
MTEB English v2 Mean(Task)75.22
C-MTEB Mean(Task)73.84
MTEB Multilingual Retrieval70.88

Qwen3 Embedding 8B specifications

Architecture and serving details for Qwen3 Embedding 8B.

PublisherQwenParameters8BLicenceApache 2.0ReleasedJune 2025

Architecture

Type: Dense dual-encoder embedding model
Backbone: Qwen3
Layers: 36
Hidden size: 4096
Attention: GQA, 32 attention heads / 8 KV heads
Pooling: Final EOS-token hidden state
Native precision: BF16

Embeddings

Context: 32K tokens
Maximum dimension: 4096
Configurable dimensions: 32-4096
MRL: Yes
Instruction aware: Yes
Languages: 100+

Serving

HexGrid endpoint: POST /v1/embeddings
vLLM: Supported, >= 0.8.5 in Qwen examples
Sentence Transformers: Supported
Transformers: >= 4.51.0
TEI: Supported
Flash Attention 2: Recommended by Qwen for acceleration and memory savings

Qwen3 Embedding 8B frequently asked questions

Common questions about deploying Qwen3 Embedding 8B on HexGrid Cloud.

How many dimensions does Qwen3-Embedding-8B produce?

The full embedding is 4096 dimensions. Qwen's Matryoshka training allows user-defined output dimensions from 32 through 4096.

Should I add an instruction to every document?

For retrieval, Qwen's examples add the task instruction to the query but embed documents directly without the query instruction.

What happens if I omit the query instruction?

Qwen reports that in most tested retrieval scenarios, omitting the instruction on the query side can reduce retrieval performance by approximately 1-5%.

Can I use smaller vectors?

Yes. HexGrid's documented embeddings endpoint exposes a dimensions field, and the model supports Matryoshka output dimensions up to 4096. Evaluate the desired dimension on your own corpus before fixing the production index size.

Can I use it commercially?

Yes. The Qwen3 Embedding family is released under Apache 2.0.

Deploy Qwen3 Embedding 8B today

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

Deploy Qwen3 Embedding 8B See all models