Beam is an open-source runtime for running serverless AI workloads, developed in the open in the beam-cloud/beta9 repository on GitHub. The project has 1,827 stars and is under active development in 2026. It gives Python developers a simple way to deploy GPU inference, sandboxes, and background jobs without managing infrastructure.
Beam pairs its open-source runtime with a hosted cloud. You describe your workload in Python (an image, a GPU type, an autoscaling policy) and Beam launches containers with sub-second cold starts thanks to its custom container runtime and scheduler. Workloads scale to zero when idle and fan out to hundreds of containers when busy. The same platform provides isolated sandboxes for executing LLM-generated code, persistent storage volumes, webhooks, and scheduled jobs. The repository contains the full stack behind the hosted product, including the container runtime, scheduler, and worker components, so the platform can be inspected and extended rather than treated as a black box. Beam positions the system for the new wave of agent workloads, where software needs fast, disposable GPU computers on demand.
The beta9 runtime is open source under the AGPL-3.0 license and free to self-host. Beam’s hosted cloud is pay-as-you-go, billed for the compute and GPU time you actually use, so the overall offer is freemium.
Beam is for AI startups and ML engineers who want to ship model inference and agent sandboxes quickly, without building Kubernetes infrastructure or paying for idle GPUs.