Overview
Qwen3-Reranker-8B puts the most relevant documents first in your search or RAG application. It is an open-weight model from Alibaba's Qwen team (Apache-2.0), supporting over 100 languages and documents up to 32K tokens. TL Engine serves it with kernels written specifically for the NVIDIA L4.
Against vLLM, a mature and widely used open-source engine, running the same model in 16-bit on the same ml.g6.xlarge, TL Engine gives:
- Faster responses: requests of 20 documents run 1.6x faster. A 32K-token document takes 6.7 s instead of 11.1 s.
- More work per instance: 1.5 to 1.7x the throughput, about 31 million tokens an hour on requests of 20 documents. Batch transform jobs need 38% fewer billed hours.
- Lower cost: 22 to 26% less per token on busy endpoints and 20 to 24% less in batch jobs, software fee included.
Ranking quality is the same as the original 16-bit model's on standard benchmarks: nDCG@10 on FiQA, SciFact and ArguAna is within 0.01, and higher on some. Scores track the original model's closely.
The benchmark scripts and raw results are public: https://github.com/transcendent-logic/tl-engine-benchmarks
Highlights
- 1.6x faster responses than vLLM 16-bit, including on 32K-token documents
- 1.5 to 1.7x the throughput and 20 to 26% lower cost per token on busy endpoints and in batch jobs, software fee included
- Open-weight (Apache-2.0), over 100 languages, network-isolated: your data stays in your AWS account
Details
Features and programs
Financing for AWS Marketplace purchases
Pricing
Free trial
Dimension | Description | Cost/host/hour |
|---|---|---|
ml.g6.xlarge Inference (Batch) Recommended | Model inference on the ml.g6.xlarge instance type, batch mode | $0.30 |
ml.g6.xlarge Inference (Real-Time) Recommended | Model inference on the ml.g6.xlarge instance type, real-time mode | $0.30 |
Vendor refund policy
No refunds. Please use the 31-day free trial to evaluate the product before paid use begins. Questions: hello@transcendentlogic.com .
How can we make this page better?
Legal
Vendor terms and conditions
Content disclaimer
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Version release notes
First release. Qwen3-Reranker-8B (Qwen's weights, unmodified) on TL Engine for the NVIDIA L4. Validated on ml.g6.xlarge for real-time endpoints and batch transform, on SageMaker's default host (NVIDIA driver 535). Maximum length 33,000 tokens for a query and one document.
Additional details
Inputs
- Summary
One JSON object per request (content type application/json):
- query: a string, the search query or question
- documents: a list of strings, the candidate passages in any order
Send plain text: the container applies the model's prompt template and tokenizes. Real-time and batch take the same format.
- Limitations for input type
- Up to 33,000 tokens for a query and one document; up to 25 MB per request, real-time or batch (at the recommended MaxPayloadInMB of 25). Batch transform takes one request per input file.
- Input MIME type
- application/json
Resources
Support
Vendor support
For support, email hello@transcendentlogic.com .
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
