At the 2025 PyTorch Conference, the PyTorch Foundation announced several initiatives aimed at advancing open, scalable AI infrastructure. The foundation welcomed Ray, a distributed computing framework, as a hosted project and also introduced PyTorch Monarch, a new framework that simplifies distributed AI workloads across multiple machines. The event also spotlighted new open research projects, including Stanford’s Marin and AI2’s Olmo-Thinking, highlighting the growing push for transparency and reproducibility in foundation model development.
The inclusion of Ray reflects the foundation’s broader strategy to build a unified open ecosystem spanning model development, serving, and distributed execution. Originally developed at UC Berkeley’s RISELab, Ray provides a compact set of Python primitives that make distributed computation as intuitive as writing local code, enabling developers to scale training, tuning, and inference workloads seamlessly.
The addition of Ray complements other recent projects under the foundation’s umbrella, including DeepSpeed for distributed training and vLLM for high-throughput inference. Together, PyTorch, DeepSpeed, vLLM, and Ray form a cohesive, open-source stack that covers the full model lifecycle—from experimentation to production-scale deployment.
In parallel, the Meta PyTorch team introduced PyTorch Monarch, a framework designed to abstract entire GPU clusters as a single logical device. Monarch’s array-like mesh interface allows developers to express parallelism using Pythonic constructs while the system automatically manages data and computation distribution. Built on a Rust-based backend, Monarch aims to combine performance with safety, thereby reducing the cognitive load associated with distributed programming.
The conference further emphasized open collaboration in foundation model development and research. In a keynote, Percy Liang from Stanford University introduced Marin, an open lab under the Center for Research on Foundation Models, which seeks to make frontier AI development fully transparent by releasing datasets, code, hyperparameters, and training logs to enable reproducibility and community participation.
Similarly, Nathan Lambert, a senior research scientist at Ai2, presented Olmo-Thinking. This is an open reasoning model that has disclosed details about the training process, model architecture decisions, data sourcing, and training code design, which are often absent in closed model releases. These efforts align with a broader movement toward open and reproducible foundation models.
By expanding its scope beyond core framework development, the PyTorch Foundation is positioning itself as a central hub for open AI infrastructure. The upcoming 2026 PyTorch Conference in San Jose is expected to continue this focus on ecosystem collaboration and developer enablement.