Flyte 2 is GA: durable distributed AI workflows using regular Python
5 hours ago
- Flyte 2 is now generally available as an open-source, durable AI runtime under the Linux Foundation, rebuilt from the ground up for modern AI and ML workloads.
- It is designed to be easy to learn: pure Python with no DSL, promises, or workflow language—tasks call tasks, and control flow is natural Python.
- Flyte 2 provides infrastructure-level durability, recovering not only from code bugs but also from compute failures like OOM crashes or preempted nodes.
- It unifies real-time and batch processing under one programming model, supporting scheduled pipelines, agent loops, and inference endpoints.
- Key features include dynamic control flow, agent-native loops with durable non-deterministic steps, orchestration sandboxes, human-in-the-loop, and MCP integration.
- Users can keep their data and infrastructure private, running on their own laptop, devbox, or Kubernetes cluster without vendor lock-in.
- Migration from Flyte 1 is supported with free consultations, a migration hub, and office hours; Flyte 2 has already had 1.3 million downloads before GA.