On Device AI
Ground local AI with instant access to user context.
Backed by
The Local Search Engine for AI.
Semantic, keyword, and hybrid search that runs entirely on device. Embed, index, and retrieve in under 10ms with no cloud infrastructure, no vector database, and no network round trip.
Your AI app
Retrieval Runtime powered by Moss
Ask a question…
<10ms
Retrieval
32 MB
One Model
100%
Offline
CPU Only
No NPU Required
Used in production by companies backed by Y-Combinator & a16z
Problem
On-device AI is here. Retrieval still needs the cloud.
Modern AI runs on phones, laptops, browsers, vehicles, wearables, and edge devices. Retrieval still runs in the cloud, adding latency, infrastructure cost, and privacy tradeoffs.
Cloud
Retrieval
Retrieval
Cloud round trip
Cloud Search Infrastructure
- Every query needs a network request
- Vector databases become part of your architecture
- Costs scale with usage
- No offline support
Traditional Embedding Pipelines
- Large models increase app size
- GPU and NPU compete with your AI workloads
- Built for servers, not edge devices
- Too heavy for phones and consumer hardware
Local AI needs a different retrieval architecture.
Retrieval that runs where your AI runs.
Moss brings semantic, keyword, and hybrid search directly to where your AI runs, eliminating cloud latency, reducing infrastructure costs, and keeping data on device.
Moss on Device AI
- faster
- lower cost
- privacy
Intelligent In-App Search
Search everything inside your application by meaning, not keywords.
Cross-App Retrieval
Retrieve context across files, notes, photos, messages, and application data.
Built for Production
Runs entirely on CPU with no cloud infrastructure or external services.
Messages1ms
100K+ msgs
SMS, chat, attachment context
Files1ms
10K+ files
Document text + filenames + metadata
team offsite san francisco
Gallery
Notes
offsite agenda — day 1, Fort Mason
Messages
…flights into SFO are booked for the offsite!
Files
Offsite plan June 20.docx
Offsite plan June 20 (2).pdf
Gallery3ms
50K+ photos
Image captions + EXIF + people tags
Notes0.8ms
100K+ notes
Notes body + handwriting OCR
The complete retrieval pipeline. Running locally.
Moss runs the entire retrieval pipeline on device.
Whether you're powering AI assistants, semantic search, RAG, recommendations, memory, or system search, it all happens locally.
No cloud APIs.
No embedding service.
No vector database.
No round trip.
Just instant retrieval where your users already are.
Moss engine
- Embedding query locally
- Searching memory index
- Running hybrid search (vector + BM25)
- Scanning 142,384 memories
- Applying RRF ranking
- Applying recency signals
- Filtering by workspace permissions
- 3 relevant memories retrieved
- last_month_investor_update
- mrr_chart · recently opened
- leads · updated yesterday
- Total latency: 4.8ms
Your AI app
finish the investor update
I found last month's update, the MRR chart, and the Leads from yesterday.
I'll draft this month's version with the same structure and update the metrics.
- Content
- EmbeddedOn Device
- IndexedOn Device
- RetrievedOn Device
- Grounded AI & Search
One SDK. Complete local retrieval.
Search
Everything developers need in one SDK.
- Semantic Search
- Keyword Search
- Hybrid Search
- Cross Index Search
Embeddings
One lightweight model for every modality.
- Text
- Images
- Audio
- 32 MB INT8
- CPU Execution
Retrieval Engine
Purpose built for edge hardware.
- Proprietary indexing
- Memory mapped retrieval
- Instant indexing
- Less than 1% recall loss after quantization
Built for modern AI products.
On Device AI
Ground local AI in documents, messages, notes, photos, and app data. Nothing leaves the device.
Consumer Apps
Search every piece of user content inside your app. Built for productivity, media, ecommerce, travel, finance, and health.
OEMs
One local retrieval engine across phones, vehicles, TVs, wearables, and embedded devices. System-wide search, fast and offline.
Enterprise Applications
Secure AI search that keeps sensitive company knowledge on employee devices.
Faster than cloud search.
| System | Average Search Latency |
|---|---|
| Moss | 3.1 ms |
| ChromaDB | 351.8 ms |
| Pinecone | 432.6 ms |
| Qdrant | 597.6 ms |
Benchmarks include embedding inference and retrieval across 100,000 documents. Most competing benchmarks measure retrieval only.
The future of AI is local.
Retrieval should be too.
Embed. Index. Retrieve. All on the device.