TinyLink

πŸ“Œ Summary

TinyLink is a lightweight language model fine-tuned to translate natural language instructions into MAVLink command dictionaries. It is designed for edge robotics. Unlike solutions relying on cloud APIs, TinyLink runs fully offline on your local machine.

image/png

Features

  • Translates plain text instructions into MAVLink commands.
  • Runs entirely on-device for enhanced privacy. No API keys or cloud dependency.
  • Runs on everyday hardware; no GPU or excessive RAM needed.
  • Tested with ArduPilot SITL.
  • Achieves 1.5–2.2s inference times on CPU, depending on hardware.
  • Supported Commands:
    1. Arm
    2. Disarm
    3. Takeoff
    4. Land
    5. Change mode (limited modes supported)
    6. Move in X, Y, Z (Copter and Rover)

Performance & Tested Platforms

Platform RAM Inference Time (avg) Status
Win 11 (App) & WSL2 (SITL) 16 GB 1.7 - 4s (Avg 2.2s) βœ… Tested
Raspberry Pi 5 4 GB 0.8 - 2s (Avg 1.5s) βœ… Tested
NVIDIA Jetson Nano - - ❌ Not tested

Demo

For a demo on how to use this model, you can check the following Github repo

Downloads last month
69
GGUF
Model size
1.1B params
Architecture
llama
Hardware compatibility
Log In to view the estimation

We're not able to determine the quantization variants.

Video Preview
loading