​
FuncMaster-v0.1-Mistral-7B

Release Date: 3/1/2024

FuncMaster-v0.1-Mistral-7B represents the first public release of our ambitious project to bring function-calling capabilities directly to your device. As an initial release, it lays the groundwork for a new class of on-device AI models, showcasing our core vision and inviting the community to explore its potential.

​
What’s New

  • Function Calling on Device: Introducing the ability to run function calls locally on your computer and iPhone. This is a significant step towards more interactive and responsive AI applications that can operate independently of cloud services.
  • Local Execution: Designed to run directly on devices, FuncMaster-v0.1-Mistral-7B marks a move towards decentralizing AI capabilities, offering users privacy, control, and reduced latency in AI interactions.
  • Initial Device Compatibility: Compatibility testing has begun, with the model running on Mac’s and iPhones, setting the stage for broader device support in future releases.

​
Improvements

  • Speed Optimization: Achieves ~9 tokens/second on an iPhone 15 Pro and ~24 tokens/second on a Mac M1. These performance benchmarks offer a glimpse into the model’s efficiency and potential for real-time applications.
  • Prototype Functionality: While still in the early stages, the model demonstrates the feasibility of on-device function calling, serving as a proof of concept for further development.

​
Known Issues and Limitations

  • Accuracy and Reliability: The model’s current ability to accurately interpret and execute function calls is a primary focus for improvement. Users may experience varying degrees of accuracy, particularly with complex or uncommon queries.
  • Performance Variability: Device performance can vary significantly, influenced by hardware specifications, system load, and the complexity of tasks. This variability is an area of ongoing optimization.
  • Energy Consumption: Preliminary feedback indicates increased energy consumption on portable devices, a consideration that is being addressed as we work to balance performance with efficiency.
  • Messages: Instruct messages aren’t correct but the function call is correct, there are also sometimes misspelling in words or incorrect formatting of the JSON.

​
Future Directions

FuncMaster-v0.1-Mistral-7B is just the beginning. Our roadmap includes enhancing model accuracy, expanding device compatibility, optimizing for energy efficiency, and broadening the range of functions and tasks the model can perform. We are committed to regular updates and improvements, driven by community feedback and our dedication to innovation.

Introduction