MetalBender – the framework from Xmartlabs’ Bender project – is an open-source Swift framework that makes it easy to craft and run fast neural networks on iOS, with Apple’s Metal GPU framework doing the heavy lifting under the hood.
Developed by Xmartlabs and released under the MIT license, Bender is an abstraction layer over MetalPerformanceShaders (MPS), Apple’s GPU compute framework. Instead of wrestling with MPS directly, developers define networks from common building blocks – convolution, pooling, fully connected layers and normalizations – and Bender translates them into efficient Metal-backed operations that run on the iPhone or iPad GPU. A standout feature is its TensorFlow adapter: models trained in TensorFlow can be loaded as frozen graphs and automatically translated into Bender layers, so teams can train on servers and deploy the same model inside an iOS app without bundling TensorFlow itself. The project has collected about 1,800 stars on GitHub; its last push was in 2023, so it is best seen as a mature, stable library from the era before Core ML matured, still valuable when you need direct control over Metal-level inference.
Bender is completely free and open source under the MIT license. There are no paid tiers, subscriptions or hosted services – you simply add the framework to your own app.
Bender suits iOS developers who want low-level, GPU-accelerated neural-network inference through Metal, especially teams maintaining apps built on MetalPerformanceShaders or porting TensorFlow models to iPhone and iPad without adopting a large runtime.