Performance comparison across vector databases for 100k dbpedia dataset
Antarys, Chroma, Qdrant and Milvus
Antarys Scales At Multimodal
Antarys is built from the ground up specific optimisations made for certain vector dimensions, enabling unparalleled performance
Check Image BenchmarksTest Configuration - 1,000 vectors, 512 dimensions, ResNet34 model, Afghan Hound query image
Popular Image Embedding Dimensions
Common embedding dimensions for popular image models and Antarys specific optimization support
| Model / Framework | Embedding Dimension | GPU Optimized | CPU Optimized |
|---|---|---|---|
| CLIP (ViT-B/32) | 512 | ||
| CLIP (ViT-B/16) | 512 | ||
| CLIP (ViT-L/14) | 768 | ||
| OpenCLIP (ViT-H/14) | 1024 | ||
| ResNet-based embeddings | 2048 (e.g., ResNet-50) | ||
| EfficientNet (e.g., B0-B7) | 1280 to 2560 | ||
| DINOv2 (ViT-G) | 1536+ | ||
| MobileNet | 1024 (varies by config) |
Vector Database Cost Analysis
Compare operational costs across vector databases based on query performance from our benchmark, this is just an estimated model
| Database | QPS | Latency | Estimated Cost | vs Cheapest |
|---|---|---|---|---|
AntarysMost Cost-Effective | 75.7 | 13.2ms | $2.0004 | Baseline |
Qdrant | 26.8 | 37.3ms | $2.0425 | +$0.0421 (1.0x) |
Milvus | 14.5 | 69.1ms | $2.098 | +$0.0977 (1.0x) |
Chroma | 9.1 | 109.9ms | $2.1694 | +$0.1691 (1.1x) |
Pinecone | 1.7 | 585.1ms | $3.0002 | +$0.9998 (1.5x) |
Built to handle any embeddings We fine tuned our vector database to specific dimensions for maximum performance
We built the database from the ground up So you can make AI work anywhere
Strategic Algorithmic Optimisation For Unbeatable Performance and Accuracy
Benchmark Summary
Comprehensive performance comparison across vector databases
| Metric | Antarys | Chroma | Pinecone | Qdrant | Milvus |
|---|---|---|---|---|---|
| Write Performance | |||||
| Throughput (vectors/sec) | 2220 vec/s | 2214 vec/s | 2507 vec/s | 743 vec/s | 1961 vec/s |
| Avg Batch Time (ms) | 450.4 ms | 451.7 ms | 39.9 ms | 1345.6 ms | 509.9 ms |
| P99 Latency (ms) | 493.2 ms | 475.4 ms | 60.0 ms | 1461.1 ms | 830.6 ms |
| Read Performance (1000 queries) | |||||
| Throughput (queries/sec) | 75.7 q/s | 9.1 q/s | 1.7 q/s | 26.8 q/s | 14.5 q/s |
| Avg Query Time (ms) | 13.2 ms | 109.9 ms | 585.1 ms | 37.3 ms | 69.1 ms |
| P99 Latency (ms) | 19.2 ms | 240.7 ms | 861.6 ms | 63.3 ms | 337.8 ms |
| Accuracy | |||||
| Recall@100 (%) | 100.00 | 100.00 | 100.00 | 100.00 | 99.96 |
| Recall Dev | 0.0000 | 0.0000 | 0.0000 | 0.0000 | 0.0028 |
Antarys was benchmarked with HTTP/1.1 while every other system used gRPC, eventual shift to gRPC will greatly increase overall performance by a factor of 1.6x (expected)
Common Questions & Answers
Find out how did we benchmark and what optimisations led to this results
Why didn't we use other open source benchmarks?
Other famous tests like qdrant's benchmark or zilliz VectorDBBench doesn't really support async python, we had to use loop.run_until_complete which made our python runtime run slow and until we have a working gRPC client and server, we will be using this simple benchmark which uses dbpedia dataset from huggingface
What hardware did we use to run the benchmarks
An Apple M1 Pro with 16GP RAM and 512GB SSD with 8 cores (6 performance and 2 efficiency)
What optimisations make Antarys compete with industry standard databases?
Antarys was built from the ground up. And not only did we reimagine how HNSW indexing works, we built an architecture around it to support our algorithm to run faster, consume less memory and yield faster responses with accuracy.
What is the A-HNSW Algorithm?
AHNSW stands for Async HNSW, the proprietary algorithm that powers our vector database. At the heart of AHNSW, we have made critical optimisations by leveraging parallel execution without thread lock, allowing much faster graph navigation while producing standard accuracy for your LLM apps
When will we release Antarys Cloud?
We are a young company working continuously to improve our vector databse engine, Antarys's vector database is a byproduct of our AHNSW and we are constantly researching on making AI faster, accessible and native, we are expecting to see Antarys Cloud by Jan 2026!
Mission of Antarys
Our mission is to make AI hardware and software accessible to everyone and everywhere, make AI deployable and run anywhere. There was a time when having computers at home was nothing short of a vision, a bold idea. We envision the same thing for AI. The heart of Antarys is to make AI run anywhere with our local first party AI infrastructre
Get Started in Minutes
Antarys can be self hosted and work offline with your existing LLM stack, get started now for your system with this one command, requires curl
curl -fsSL http://antarys.ai/start.sh | bash- Antarys Database
- Vertically Scaled by default
- Supports every feature
- Best for your hobby LLM project
- Antarys Database
- Vertically Scaled by default
- Supports every feature
- Best for your hobby LLM project
Apple ARM
Apple binaries for Intel and ARM mac for everybody
Free for offline forever.
- Antarys Database
- Vertically Scaled by default
- Supports every feature
- Best for your hobby LLM project
- Antarys Database
- Vertically Scaled by default
- Supports every feature
- Best for your hobby LLM project