community
Research code and libraries that use Raven, from the people who write them.
Built something with Raven? Add it to this page with a pull request, or tell us about it in an issue.
01 — research
SOFO, a second-order optimizer that never backpropagates. Each step measures the loss and its curvature in a random subspace of the parameters with one batched forward-mode pass, so its memory does not grow with the length of a sequence.
SYMO estimates a network's curvature from gradients averaged over the symmetries of its weights, and preconditions gradient descent with the estimate.
Artemev, Xia, Boyd, Yu, Dangel, Hennequin and Bernacchia, ICML 2026 ↗
iLQR-VAE learns input-driven dynamics from neural recordings. Its encoder is an optimal-control solve: iLQR infers the inputs that best explain each trial, and the model is fit end to end by differentiating through that solve.
The Hennequin lab also teaches Raven to its new members: its getting-started course ↗ works through Nx, Rune and Vega with runnable examples and exercises, ending in a small model trained end to end.
02 — libraries
Classical machine learning, like scikit-learn: immutable estimator specifications, pipelines fitted only on training rows, deterministic cross-validation and search, and portable fitted models. Its modelkit-nx and modelkit-talon packages take Nx tensors and Talon columns, checked.
Batched geometric algebra for robotics: Euclidean, projective and conformal models of space, with forward and differential kinematics. Every multivector is an Nx tensor, so it broadcasts, differentiates with Rune and compiles.
Differentiable iterative LQR for discrete-time optimal control. Gradients come implicitly from the solution rather than by unrolling the iterations, in reverse or forward mode, and a fixed-iteration solve compiles with Rune into one program whose size does not grow with the horizon.
Audio processing for OCaml, on Nx: reading, writing and resampling audio files.
in use
We chose OCaml early, for the language itself: the type system, the expressiveness, the confidence that comes from code that doesn’t fall over at runtime. For years we made that choice despite the numerical ecosystem, not because of it. We wrote a lot of our own tooling, and we accepted the tradeoff.
Raven removes the tradeoff. Our models are type-safe from end to end. JIT compilation to GPU makes acceleration seamless, without rewriting our code or leaving the language. And the effects-based design is a joy to use: it feels like OCaml, not like a port of something else.
What impressed me most was how effortless raven made porting and unifying our previously disparate libraries under one framework. We have never been more productive.”