Qwen3-Embedding-8B: specs, benchmarks, and how to run it on Lyceum

Get started: call Qwen3-Embedding-8B on Lyceum

Lyceum provides an OpenAI-compatible API for Qwen3-Embedding-8B. Because the endpoint mirrors the standard OpenAI specification, you can switch your embedding provider by updating the base URL and API key, requiring minimal changes to your downstream RAG logic.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.embeddings.create(
 model="Qwen/Qwen3-Embedding-8B",
 input="Your text to embed",
 )

Pricing and region for Qwen3-Embedding-8B

This model is served through Serverless Inference, which is billed per token with no base fee and has no service tiers: the model you address sets the balance between latency and capability. It is hosted in the eu-north1 region, so requests for it are processed in European data centres in Spain, Paris and the Nordics.

The pricing for Qwen3-Embedding-8B is $0.01 per million tokens. Because this is an embedding model, billing applies exclusively to input tokens; there are no output tokens generated. Lyceum charges no egress fees, allowing you to transfer large volumes of vector data to your vector database without incurring hidden network costs.

What Qwen3-Embedding-8B is good at

Multilingual retrieval and cross-lingual search

Qwen3-Embedding-8B supports over 100 natural and programming languages. It excels in cross-lingual retrieval tasks where the search query and the target document are in different languages. This makes it highly effective for global enterprise search systems and multilingual RAG pipelines, outperforming previous-generation models on MTEB (Multilingual), the massive multilingual text embedding benchmark.

Instruction-aware embedding generation

The model architecture is instruction-aware, meaning it can adapt its vector representations based on task-specific prompts. By prepending an instruction (e.g., Instruct: Given a web search query, retrieve relevant passages that answer the query.\nQuery: {text}), developers can optimize the embeddings for specific downstream tasks like classification, clustering, or asymmetric retrieval. The Qwen team reports that using tailored instructions yields a 1% to 5% performance improvement across various benchmarks.

Matryoshka Representation Learning (MRL)

While the model natively outputs 4,096-dimensional vectors, it was trained using Matryoshka Representation Learning. This allows developers to truncate the output vectors to lower dimensions (e.g., 1024 or 256) and re-normalize them without a catastrophic drop in retrieval accuracy. This flexibility is critical for teams managing vector database storage costs, as it permits a trade-off between storage footprint and semantic precision.

Limitations and trade-offs

Where Qwen3-Embedding-8B falls short

As an 8-billion parameter model, Qwen3-Embedding-8B requires significant VRAM for deployment compared to standard embedding models. While Lyceum abstracts this infrastructure complexity through our Serverless Inference API, teams attempting to self-host the model will find it demands substantial GPU resources, typically requiring an NVIDIA A100 or H100 for high-throughput concurrent requests.

Additionally, while the model supports a 32,768-token context window, processing documents at this maximum length incurs high computational overhead. The quadratic scaling of attention mechanisms means that embedding massive documents in a single pass will result in higher latency. For production RAG systems, chunking documents into smaller segments (e.g., 512 to 2,048 tokens) remains the recommended approach to maintain low latency and ensure the highest retrieval accuracy, as embedding models generally capture semantic density better in shorter passages.

Finally, the model is a text-only embedding system. It does not support multimodal inputs. Teams requiring image or video embeddings must use a dedicated multimodal model, such as the Qwen3-VL-Embedding variants, rather than this text-specific architecture.

Benchmarks and how it compares

Qwen3-Embedding-8B benchmark results

Qwen3-Embedding-8B achieves state-of-the-art performance on the Massive Text Embedding Benchmark (MTEB), particularly in multilingual and retrieval-heavy subsets. The table below highlights its performance against both its smaller siblings and other prominent embedding models.

Scores are MTEB (Multilingual) results from the Qwen3-Embedding-8B model card, also published in the Qwen3-Embedding GitHub repository, read 3 August 2026. Qwen notes that the scores for compared models were taken from the MTEB leaderboard on 24 May 2025.

Comparison to sibling models

Within the Qwen3 embedding family, the 8B model offers the highest semantic accuracy, scoring 70.58 on the MTEB mean compared to the 4B model's 69.45 and the 0.6B model's 64.33. The 8B variant is the optimal choice for complex enterprise RAG pipelines where retrieval precision is paramount. However, for latency-constrained applications or teams processing billions of tokens where cost is the primary driver, the Qwen3-Embedding-0.6B provides a highly efficient alternative, trading a few percentage points of accuracy for significantly lower computational overhead.

Using it in production

Production configuration for Qwen3-Embedding-8B

When integrating Qwen3-Embedding-8B into a production environment, proper configuration of the API request ensures optimal performance. The model is accessed via the standard embeddings.create endpoint, which is the embeddings route of the OpenAI-compatible API rather than chat completions. Serverless Inference has no tier to select: you address the model by name and pay per token, and requests are served from the eu-north1 region, which suits both bulk document ingestion and real-time query embedding.

To maximize retrieval quality, prepend task-specific instructions to your queries. For example, when embedding a user's search query, format the input string as Instruct: Given a web search query, retrieve relevant passages that answer the query.\nQuery: {user_input}. When embedding the documents themselves for storage in your vector database, instructions are typically omitted, allowing the model to generate a neutral semantic representation of the text.

Calculating per-token pricing

Lyceum bills Qwen3-Embedding-8B at $0.01 per million tokens for input processing. Because embedding models do not generate text, there are no output token costs.

For a realistic production workload, consider a pipeline that ingests 50,000 documents, each averaging 800 tokens. The total input volume is 40 million tokens. At $0.01 per million tokens, the total compute cost to embed this entire dataset is $0.40. Furthermore, because Lyceum does not charge egress fees, transferring the resulting 4,096-dimensional vectors to your external vector database incurs zero additional network costs, making this predictable for large-scale enterprise deployments.

Running Qwen3-Embedding-8B on EU-sovereign infrastructure

Why run Qwen3-Embedding-8B on Lyceum

For European enterprises and AI startups, data residency is a regulatory requirement. Processing sensitive corporate documents, medical records, or proprietary code through US-hosted APIs introduces compliance considerations. Lyceum addresses these requirements by hosting Qwen3-Embedding-8B on EU-sovereign infrastructure in our eu-north1 region.

Requests to this model are served through Serverless Inference from European data centres in Spain, Paris and the Nordics, with GDPR-compliant processing, no training on customer data, and no retention of inference prompts or outputs. Serverless Inference is billed per token with no base fee, and the published rate for Qwen3-Embedding-8B is $0.01 per million input tokens on the Lyceum pricing page.

Furthermore, Lyceum is built on open-stack transparency. We utilize optimized open-source inference engines like vLLM and NVIDIA Dynamo rather than proprietary black-box systems. This supports customer portability; you are not locked into a proprietary ecosystem. With our drop-in OpenAI-compatible API, migrating your RAG application to GDPR-compliant LLM inference in Europe requires only a change to the base URL, allowing your engineering team to focus on building features rather than managing infrastructure.

Sources

[1] Qwen/Qwen3-Embedding-8B - Hugging Face; [2] QwenLM/Qwen3-Embedding - GitHub; [3] Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models - arXiv; [4] qwen3-embedding:8b - Ollama; [5] Qwen3-Embedding - a Qwen Collection - Hugging Face