Roadmap

Raven is in pre-alpha. We're building towards our first milestone: an end-to-end demo of training MNIST with visualizations in a Quill notebook.

Alpha Release

Our immediate goal is demonstrating the complete ML workflow in OCaml:

• Train a neural network on MNIST using Kaun
• Visualize training metrics with Hugin
• Run everything in a Quill notebook
• Achieve reasonable performance on CPU

Once we hit this milestone, we'll stabilize the APIs and cut v1.0.

Post-Alpha: Training Large Models

After v1.0, we'll shift the goalpost to training large models. Our focus will become mostly performance, as the requirement for real-world ML workflows.

JIT Compilation (!!)

• Build JIT compiler for Rune
• Target LLVM, CUDA, and Metal for hardware acceleration

Accelerated Backends

• Metal backend for Nx (macOS GPU support)
• CUDA backend for Nx (NVIDIA GPU support)
• Seamless integration with the JIT

Deep Learning at Scale

• Complete Kaun with all features needed for modern LLMs
• Train and run inference for Llama 3 as our benchmark
• Enable distributed training across multiple GPUs

Beyond

Future development depends on community adoption and potential sponsorship. If we achieve sustainable development, priorities include:

Performance Parity with Python

• Match NumPy/PyTorch performance through JIT optimization
• Prove OCaml can compete on raw compute, not just developer experience

Expanded Ecosystem (based on community feedback)

• Dataframe manipulation library
• Domain-specific libraries for common workflows

Distributed Computing

• First-class distributed training support
• Tools for deploying models at scale

But first, we need to prove the concept works. That starts with MNIST in a notebook.