Manifold Platform

Manifold is the first distributed machine learning platform that leverages functional programming.

We built a system that leaps over giant engineering obstacles, allowing companies to access ML training and deployment capabilities previously out of their reach.

Our compiler builds massive distributed systems for complex reinforcement learning and facilitates seamless edge deployment.

The Manifold programming platform is the missing link for teams that need to scale complex machine learning workflows.

It allows them to iterate on learning processes rather than troubleshooting and maintaining infrastructure.
‍
The platform allows engineers to describe their entire workflow with one high level language. The compiler deploys all of the moving parts in a robust, hardware accelerated, distributed choreography.
‍
Manifold siloes components of existing platforms, distributes work automatically and allows teams to reuse the same code for training and deployment.

Features and Benefits

Type Safe and Purely Functional

Manifold inherits all of the benefits of strongly typed purely functional programming from Haskell and Futhark. We then add on an elegant type system for safe distributed computing.

Seamless Scaling and Distribution

Scaling is an abstraction in Manifold that is determined at compile time. This allows teams to start small and scale up without design changes. They can focus on the what instead of the how.

Integrated Synthetic Data

Synthetic data is the mainstay for training specialized systems for manufacturing, robotics, bio-medical and many more fields. Manifold makes the production and integration of this data a first class citizen that can be streamed concurrently at a massive scale and integrated seamlessly with training.

Generated GPU Code

Most GPU accelerated projects are plagued by the sheer number of simple but specific boilerplate functions that must be hardware accelerated. Manifold automates the production of these functions and allows the programmer to easily embed custom GPU kernels in their workflow.

Deadlock Free by Design

It is impossible to write a program in our language that creates a distributed communication dead-lock. This means that an entire class of (really nasty) bugs are eliminated.

PyTorch Compatibility

PyTorch model architectures are embedded in Manifold for training and inference. This allows ML engineers to migrate existing work and continue to access the larger ecosystem.

Seamless Model Re-use

Trained models and other artifacts can be cached and recalled simply by calling the function that created them with the same hyper-parameters, improving an ML team's ability to find solutions quickly via the scientific method.

Pivotal Code Re-use

Manifold will compile the same interface code for neural networks onto different platforms between training and deployment allowing developers to avoid subtle mismatch bugs and the reimplementation tax.

Manifold is available for private testing with partner teams.
Global launch Q2 2025