IP Available for licensing and acquisition

The world's fastest
vector search.

MRHI is a proprietary vector index delivering sub-millisecond k-NN across CPU, CUDA, and Metal. Up to 3,156× faster builds and 1M+ QPS on published benchmarks — quietly outclassing every public vector database.

1.33M QPS
CUDA · 10k · 128d
5M/sec
Inserts
99.94%
Recall floor
3,156×
faster vs pgvector
search at scale RAG that's not slow real-time search recommendation engines agent memory semantic dedup image similarity anomaly detection search at scale RAG that's not slow real-time search recommendation engines agent memory semantic dedup image similarity anomaly detection
Benchmarks · ann-benchmarks

Numbers you can't argue with.

Public benchmark results on the standard glove-100-angular dataset. We don't cherry-pick — recall is held at 99.94%+ across all measurements.

Query throughput

queries per second · higher is better
MRHI · CUDA
1.33M
MRHI
169.9k
voyager
5.5k
weaviate
5.5k
milvus-hnsw
4.4k
opensearch
2.1k
lancedb
1.4k
pgvector
996
qdrant
111
3,156×
Faster build throughput than pgvector on the public 1M+ CPU comparison. Ingest is no longer a bottleneck.
100M
Vectors benchmarked in our published large-scale runs. Same engine, same recall, same hardware envelope.
0.001ms
p99 latency at 10k vectors on CUDA.
ROI Calculator

What you'd save with MRHI.

Plug in your current vector-search stack. Same hardware. We'll estimate the fleet you'd replace and the dollars you'd reclaim every year.

Servers running vector search
Cost per server / month $
MRHI deployment
Market capture % of $3.38B 2026 vector-search market
Estimated annual savings
$0
0% lower infrastructure cost

Estimates based on published per-server throughput. For exact numbers on your own data, request a private benchmark.

Capabilities

Three engines. One index.

MRHI runs natively on CPU, CUDA, and Metal — the same index, the same recall, picked per workload. The internals stay ours; the speed becomes yours.

CPU

Server-class CPU, fully utilized.

150,000+ QPS on commodity x86 with 99.94% recall. No GPU required, no warm-up, no cold-start tax. The default tier for most production workloads.

CUDA

NVIDIA-accelerated throughput.

Over 1 million QPS on a single GPU. Batches thousands of queries in flight with consistent sub-millisecond p99. The tier of choice when latency at scale is the product.

Metal

Apple Silicon — first-class.

Native Metal execution on M-series. Run large vector indexes on the laptop in your bag, on-device, with no network round-trip. Ideal for private and edge deployments.

#f3
0.978
#20
0.961
#e1
0.944
#02
0.927
#a8
0.911
Input
Vectors
f16 · f32 · binary
Dimensions
Tested 64 → 4096
Metric
L2 · Cosine · Dot · Hamming
Execution engines
Engine · 01
CPU — x86 · ARM
Engine · 02
CUDA — NVIDIA H100 / A100 / L40
Engine · 03
Metal — Apple M-series
Guarantee
99.94–100% recall on every engine
Output
Latency
Sub-millisecond p99
Scale
Benchmarked to 100M vectors
Deployment
Embedded · server · on-device
Use cases

Speed unlocks new shapes of search.

When a single query costs microseconds, you can re-rank, re-query, and reason in the loop — things that were impossible at 50ms.

Agent memory

Persistent long-term memory for agents. Recall conversations, tool results, and observations at full speed — even across millions of past interactions.

Recommendation

Realtime user → item nearest-neighbor lookups at scale.

Semantic search

Replace BM25 + ES for text, image, and multi-modal retrieval.

Anomaly detection

Stream embeddings → flag outliers in real time on production traffic.

Licensing

Acquire the engine. Own the advantage.

MRHI is not a SaaS. It is proprietary IP, available to a small number of enterprises via commercial license or full acquisition.

Evaluation
Private benchmark

Send us your dataset under NDA. We run MRHI against it live, on a video call, and walk you through the numbers. No binary leaves our hands.

  • NDA-gated process
  • Your dataset, our hardware
  • Live walkthrough over video
  • Same-week scheduling
Request a benchmark
Enterprise License
Custom

Perpetual or term license to deploy MRHI inside your stack. CPU, CUDA, and Metal engines included.

  • Binary distribution for your platforms
  • All three engines (CPU · CUDA · Metal)
  • Self-hosted, embedded, or on-device
  • Commercial use across your products
Inquire
Limited engagements · select partners only

See it. Then license it.

Try the public demo, then book a private benchmark on your own data. Acquisition and exclusive-license conversations are welcome.