Flyte

Flyte is an open source orchestration runtime for AI and machine learning workloads. The current generation, Flyte 2, is positioned as a durable runtime: it records the inputs and outputs of every step, so a crashed run resumes where it stopped instead of restarting, and it recovers from infrastructure failures rather than only from code exceptions. It suits platform and machine learning engineering teams running training, fine tuning, pretraining, batch inference, reinforcement learning and agent loops on their own infrastructure. Workflows are written as plain async Python functions with no separate domain language or YAML. A task environment declares images, resources, secrets, retry strategy and timeouts, while branching, loops and fan out use ordinary control flow and input and output between tasks is type checked. Because the runtime is infrastructure aware, it resubmits work after out of memory kills, preempted spot nodes or a GPU that disappears mid run, resuming from the last checkpoint. The same Python API also serves long running applications, with native integrations for vLLM, SGLang, Ollama, FastAPI and Streamlit, model endpoints that speak the OpenAI protocol, and scaling to zero when idle. Software development kits are published for Python, TypeScript, Rust, Go and Mojo, and the same code runs locally, in a development box or on a cluster. Flyte is licensed under Apache 2.0 and governed in the open under the Linux Foundation AI and Data umbrella, so self hosting the project carries no license fee. There is no pricing page. The commercial route is an enterprise platform delivered by Union.ai, whose pricing is not published and is arranged with the vendor. Teams on Flyte 1 should budget for a migration, since the vendor publishes separate migration resources.