Modular Funds Six Community Projects in Mojo and MAX Grant Program

A call for proposals, and a limited number of grants

According to Modular, it opened a call for community project proposals earlier this year. The goal was to offer small grants to Mojo and MAX libraries that form the foundation of an ecosystem. The company says it heard from a wide range of contributors and received far more strong proposals than it is able to support.

The six selected projects cover reinforcement learning, GPU physics simulation, signal processing, data formats, networking and animation. Several of the maintainers have presented their work at Modular community meetings, and more presentations are scheduled for the October and November meetings.

Reinforcement learning and GPU simulation

noeira, by Denis Laboureyras, is an open-source reinforcement learning framework written entirely in Mojo. It was created in early 2026 and currently supports 40+ RL algorithms, GPU-accelerated deep RL, custom 2D/3D physics engines, native arcade game engines and SDL3 rendering. The project grew out of frustration with performance cliffs in the Python reinforcement learning stack. Its long-term aim is an end-to-end framework in which environments, agents and training all run on the GPU in pure Mojo. In the near term it plans to publish benchmarks against Gymnasium, Stable Baselines3 and PyTorch.

bajo, by Jean-Gabriel Simard, is a parallel physics simulator designed to run thousands of RL and robotics training environments at once on a single GPU. It has already shipped a state of the art GPU sorting implementation, using Morton encoding and BVH construction, as its compute foundation. The six-month plan is to package that sorting layer and implement performant parallel collision detection and rigid-body dynamics. The result is to be scaled to a working multi-thousand-environment prototype and benchmarked against NVIDIA's Warp and Madrona.

Core libraries: FFT, JSON and HTTP

Martin Vuyk Loperena's Fast Fourier Transform implementation began at a hackathon in mid-2025. According to Modular, it is already approaching state of the art performance on both CPU and GPU, competitive with FFTW and cuFFT respectively. Near-term goals include half-sized transform support (hC2R, R2hC) and outright performance wins on the R2R, C2R and R2C transforms it already supports. Modular says that if the remaining gap can be closed, MAX's cuFFT dependency could be completely removed.

EmberJson, by Brian Grenier, is described as the most widely used JSON parsing library in Mojo today. It relies heavily on Mojo's reflection capabilities and supports both structured and unstructured parsing. The roadmap points toward a general serialization framework in the spirit of Rust's Serde, along with continued Python interoperability work. Early benchmarks already show it out-pacing Python's built-in JSON handling despite interop overhead, according to the post.

floki, by Mikhail Tavarez, is a Mojo HTTP client built on libcurl bindings. It already supports many features familiar to users of Python's requests module. Development now focuses on ergonomic access to advanced libcurl configuration and on simplifying the implementation of the underlying bindings.

Animation and what the cohort signals

momanim, by Josiah Laivin, is a 2D animation framework for Mojo inspired by Manim. It aims to use CPU and GPU acceleration to dramatically reduce rendering times, and targets an API close enough to Manim's that LLM-assisted Python-to-Mojo conversion is realistic. The project is in its infancy: basic image and video APIs and shape rendering are in place, and a 3D rendering API is next on the roadmap.

Modular describes the group as a snapshot of where its community is headed as Mojo and MAX continue to develop, mixing foundational infrastructure with applications. For developers, the list shows which libraries are being funded and what benchmarks and features to expect next. Several projects plan comparisons against established tools such as PyTorch, FFTW, cuFFT and Warp, so their results should show how far pure Mojo implementations can go.