The complete on-device voice + vision engine for smart glasses. 15 intents. Entity extraction. Hindi + English. Sub-10ms. Fits on any ARM chip.
Stage 1 runs continuously at ultra-low power. Stage 2 activates only when needed. Total on-device footprint: 55MB.
Navigate, call, message, capture, music, fitness, weather, alarms, shopping, translate, notifications, focus mode, identify, read text, assistant.
"Call Mom" extracts contact="Mom". "Navigate to Khan Market" extracts location. Contacts, times, durations, languages, frame types.
On-device neural OCR engine. Read signs, menus, prescriptions, labels through the glasses camera. No cloud processing.
Native Hinglish code-mixing. "Mujhe directions do", "papa ko call karo" classified correctly. 99 languages supported.
Every computation on-device. Zero data leaves the glasses. No cloud. No internet. Works in airplane mode, rural areas, everywhere.
Configurable wake word per brand. Rust SDK, ARM-native, ONNX export for any SoC. Drop-in integration.
| Feature | SLM360 LensNano | Ray-Ban Meta | Xreal | Google Glass |
|---|---|---|---|---|
| Latency | <10ms | 2-3s | 2-5s | 1-3s |
| Works Offline | Yes | No | No | No |
| Hindi | Yes | No | No | Limited |
| Privacy | On-device | Cloud | Cloud | Cloud |
| Total Size | 55MB | Cloud | Cloud | Cloud |
| Cost/Query | ₹0 | ₹0.80-4 | ₹1.60 | ₹0.80 |
| OCR | On-device | Cloud | No | Cloud |
| Entity Extraction | Yes | Limited | No | Limited |
| Wake Word | Configurable | "Hey Meta" | Fixed | "OK Glass" |
| Languages | 99 | English | English | 4 |
Type a command or use the mic. The model runs on-device in a single binary.
The glasses camera reads text in the real world — signs, menus, prescriptions, labels — then classifies what the user wants to do with it. Entirely on-device.
The model runs on the chip. No API calls. No per-query fees.
LensNano is built on SLM360X — our most advanced on-device AI engine. A ground-up rewrite of the entire inference stack: custom transformer architecture, GPU-accelerated training, proprietary tokenization, neural OCR, and edge deployment — all in a single unified framework.
Where it started. Hand-crafted tensor operations, byte-level tokenization, custom backward pass. Built from scratch in Rust with zero external dependencies. Proved that sub-millisecond on-device NLU was possible.
Everything rebuilt from the ground up. GPU-accelerated tensor engine with Metal and CUDA support. Proprietary BPE tokenizer. Neural OCR for vision. ONNX-based edge deployment for any ARM or RISC-V SoC. The engine that powers LensNano.