The Elixir ecosystem just got a serious machine learning upgrade. The team behind Livebook has released Bumblebee, a library that brings pre-trained Hugging Face models—think GPT2, Stable Diffusion, and more—directly into Elixir applications. Developers can scaffold neural network tasks in three clicks using Livebook’s new Smart cells, then embed those models into Phoenix web apps, Broadway data pipelines, or even Nerves embedded systems. No third-party Python dependencies required.
This didn’t happen overnight. The foundation was laid nearly two years ago with Numerical Elixir (Nx), a project that implements multi-dimensional tensors and a numerical subset of Elixir that compiles to CPU and GPU. Rather than reinventing the wheel, Nx leans on Google XLA and Libtorch bindings for compilation. Axon followed as a functional neural network library inspired by Flax and PyTorch Ignite, and Explorer brought dplyr and Polars-style dataframes to the language.
The pitch here is straightforward: Elixir runs on the Erlang Virtual Machine, which has concurrency and distribution baked in. That means developers can serve models inside existing web applications without spinning up separate Python microservices. In all scenarios, the team says Bumblebee models compile to both CPU and GPU.
What’s next? The team is shifting focus toward training and transfer learning, which would let businesses fine-tune pre-trained models for their own needs—something that’s still largely the domain of Python today. They’re also working on traditional machine learning algorithms and bringing Livebook to more platforms. Pull requests that add new model architectures to Bumblebee are explicitly welcomed. For a language that many wrote off as a web-development curiosity, Elixir is quietly building a credible ML story.
