Tiny language models
for tiny devices
On-device AI optimized for headphones, wearables, AR glasses, and low-power hardware. Zero cloud. Zero latency. Zero hallucination.
Example Usage
User command:
"turn noise cancelling to high"
Bijou output:
{
"function": "set_anc_mode",
"mode": "high"
}What's included
Synthetic Dataset Generation
Generate command → function-call pairs using teacher models for device-specific training.
Schema-Constrained Decoding
Ensures every output is valid, typed, and deterministic. No hallucinations.
Quantized Inference
Export to int8/int4 for fast local inference on resource-constrained devices.
Tool-Schema Definitions
Define the full list of actions your target device supports with JSON schemas.
Small-Model Fine-Tuning
Train 1-4B parameter models to output structured JSON only for your use case.
Browser Demo
Test the full flow: mic → STT → model → JSON → simulated device actions.