Missed TensorFlow Dev Summit? Check out the video playlist. Watch recordings

TensorFlow Lite 2019 Roadmap

Updated: August 29, 2019

The following represents a high level overview of our 2019 plan. You should be conscious that this roadmap may change at anytime relative to a range of factors and the order below does not reflect any type of priority. As a matter of principle, we typically prioritize issues that the majority of our users are asking for and so this list fundamentally reflects that.

We break our roadmap into four key segments: usability, performance, optimization and portability. We strongly encourage you to comment on our roadmap and provide us feedback in the TF Lite discussion groups and forums.


  • New model converter
    • New MLIR-based TensorFlow Lite convertor that better handles graph conversion (e.g., control flow, conditionals, etc...)
    • Improved diagnostics and debugging of model conversion failures
  • Expanded ops coverage
    • Prioritized op additions based on user feedback
  • Improvements to using TensorFlow ops in TensorFlow Lite
    • Pre-built libraries available via Bintray (Android) and Cocoapods (iOS)
    • Smaller binary size when using select TF ops via op stripping
  • LSTM / RNN support
    • Full support of conversion for LSTMs and RNNs
  • Pre-and-post processing support
    • New support library for model-specific pre-and-post processing
    • Utilities for common platform-specific functionality, e.g., loading a model efficiently from assets, or converting a Bitmap to a tensor
  • Control Flow & Training on-device
    • Support for control flow related ops
    • Support for training on-device, focused on personalization and transfer learning
  • Graph visualization tooling
    • Provide enhanced graph visualization tooling
  • More models and examples
    • More models on the support section of the site
    • Additional examples to demonstrate model usage as well as new features and APIs, covering different platforms.
    • Model customization libraries and tutorials to let beginners to customize those models easily.


  • Better tooling
    • Simpler benchmarking and profiling tools for understanding available accelerators and performance tradeoffs
    • Public dashboard for tracking performance gains with each release
  • Improved CPU performance
    • Continued optimization of float and quantized kernels
    • First-class x86 support
  • Updated NN API support
    • Full support for new Android Q NN API features, ops and types
  • GPU backend optimizations
    • OpenCL and Vulkan support on Android
    • Metal and Objective-C CocoaPods for Metal acceleration
  • Hexagon DSP backend
    • Initial release of DSP acceleration for pre-Android P devices


  • Quantization
    • Post training quantization for hybrid kernels -- Launched
    • Post training quantization for (8b) fixed-point kernels -- Launched
    • Training with quantization for (8b) fixed-point kernels
    • Extend post and during training APIs to (8b) fixed-point RNNs
    • Training with quantization for low bit-width (< 8b) fixed-point kernels
  • Pruning / sparsity
    • Magnitude based weight pruning during training -- Launched
    • Support for sparse model execution


  • Microcontroller Support
    • Add support for a range of 32-bit MCU architecture use cases for Speech and Image Classification
    • Sample code and models for vision and audio data
    • Full TF Lite op support on microcontrollers
    • Support for more platforms, including CircuitPython support