Flyte
flyte.org
- Free plan
- Open source
- Linux
- Web
- API
- Self-hosted
What Flyte is
Flyte 2 is a runtime for workloads involving AI training, serving, and agents. Users define workflows as Python async functions and use a TaskEnvironment to specify images, resources, secrets, and retry policies. Workflows can branch, loop, and make decisions during execution. If a task fails, Flyte can resume from that point and recover from infrastructure failures such as out-of-memory kills, preempted nodes, or disappearing GPUs. For serving, its native app integrations include Streamlit, FastAPI, vLLM, SGLang, and Ollama, with replicas able to scale to zero while idle. Other listed integrations include Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake, Weights & Biases, OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, MLflow, Dask, and Databricks. Flyte can run locally with a devbox or on a user's cluster. Union.ai says its data plane runs in the customer's cloud, and workflow inputs, outputs, logs, code, and reports do not pass through its control plane. Flyte OSS is free and open source under Apache 2.0.
Who it suits
Flyte suits developers and teams building AI workflows for training, serving, or agents. Teams can run it locally or on their own cluster, with paid Team and Enterprise plans also listed.
Good to know, and what to watch
- Workflows use Python async functions.
- Failed tasks can resume and recover from infrastructure failures.
- Execution supports branching, loops, and runtime decisions.
- Serving integrations include Streamlit, FastAPI, and vLLM.
- The OSS runtime is free and can run on a user's cluster.
- Team plan costs $950 a month plus usage.
- Team plan has a 1,000 concurrent action limit.
- Team plan includes one cluster.
- Enterprise pricing is custom and not listed.
Plans
Every plan and priceEnterprise
Flyte OSS
Team
Union.ai Team (enterprise Flyte platform)
Free plan
Getting started
- Visit https://flyte.org/.
- Install with `pip install flyte`.
- Start a local devbox with `flyte start devbox`, or run Flyte on your cluster.
- Choose the free Flyte OSS runtime, Team, or Enterprise plan based on deployment and support needs.
Limits to check
The free OSS plan is described as an open-source runtime that runs locally or on your cluster. Team lists a $950 monthly minimum credited to usage, one cluster, 30-day data retention, 1,000 concurrent actions, and a 25,000-action-per-run fanout limit. Enterprise lists custom concurrent actions, custom retention, and 3+ clusters, with custom pricing.
The Club's review
Overview
Flyte is a Python-based runtime for coordinating AI workloads that may need to recover from failure or make decisions as they run. It is best suited to teams building training, serving, or agent workflows who can operate a local environment or cluster. Its main trade-off is the cost of the managed Team plan: $950 per month before usage.
Key features
Flyte 2 uses ordinary Python async functions to define workflows. A TaskEnvironment declares each task’s images, resources, secrets, and retry policy, giving developers a direct way to describe what execution needs alongside workflow logic. That approach will suit Python teams; the stated authoring model offers less reason to choose Flyte if a team wants to build workflows without writing Python.
Recovery is a meaningful strength for work that is expensive or slow to repeat. Flyte can resume a failed task and recover from infrastructure problems such as out-of-memory kills, preempted nodes, and disappearing GPUs. Automatic retries and event triggers support that operational model, while execution observability helps teams follow runs.
Workflows can branch, loop, and make runtime decisions, which supports agent tasks whose next action depends on earlier results. Native integrations cover Streamlit, FastAPI, vLLM, SGLang, and Ollama; replicas can scale to zero when idle. The wider integration set includes Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake, Weights & Biases, OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, MLflow, Dask, and Databricks. This breadth makes Flyte a plausible fit for teams connecting orchestration to existing data and ML tools, though the integration names alone do not establish how much setup any one connection requires.
Pricing
Flyte OSS costs nothing under the Apache 2.0 license. It runs locally or on a customer’s cluster, making it the practical starting point for individuals and teams willing to manage deployment themselves. A free trial is also offered for Team.
Team costs $950 a month, billed at $950 per month plus usage. The $950 monthly minimum is credited to usage, and the 30-day free trial gives teams a month to evaluate it. The plan includes 1,000 concurrent actions, one cluster, 30-day data retention, a 25,000 action-per-run fanout limit, and one-business-day Sev 1 support. Those limits suit a team with a defined cluster footprint and substantial concurrent execution, but the monthly minimum is a steep commitment for smaller projects.
Enterprise has custom pricing, with custom concurrent-action limits and retention, three or more clusters, and white-glove support. It is aimed at organizations whose scale or support needs exceed Team’s fixed limits. The OSS runtime avoids a subscription fee, but puts cluster operation in the customer’s hands; Team adds a paid service tier with stated capacity and support terms.
Platforms
Flyte is available through an API and web interface, and supports Linux, self-hosted, and local deployment. Users can install it with pip install flyte and start a local devbox with flyte start devbox, or run it on their own cluster. Union.ai says the data plane runs inside the customer’s cloud, with workflow inputs, outputs, logs, code, and reports kept out of its control plane. Its security page lists SOC 2 Type II, SOC 2 Type I, and HIPAA certifications. This deployment model may appeal to organizations that need workloads to run in their cloud, while requiring a willingness to manage cluster infrastructure.
Who it's for
Flyte fits Python teams running durable AI workflows, especially when infrastructure interruptions, dynamic execution, or connections to serving and ML tools matter. Its local and self-hosted options also suit teams that want to run the runtime on their own infrastructure. Teams seeking a low-cost managed entry point may find the Team minimum hard to justify, while groups unwilling to operate a cluster may not get the benefit of the free OSS option.
Pros and cons
- Pro: Resuming tasks after infrastructure failures can avoid restarting long-running AI work from the beginning.
- Pro: Python async authoring and runtime branching support both conventional workflows and agent execution patterns.
- Pro: Local and customer-cluster deployment, combined with the stated data-plane arrangement, gives infrastructure-conscious teams control over where workloads run.
- Con: Team starts at $950 a month plus usage, a material cost for small teams or early experiments.
- Con: Team is capped at one cluster, 1,000 concurrent actions, and 25,000 actions per run, so larger deployments need Enterprise.
- Con: The free OSS runtime requires users to run it locally or on their own cluster, placing deployment responsibility on them.
Alternatives
Choose Kestra if you want an Apache 2.0 platform with support for a wider set of listed operating systems, including macOS and Windows. DataKitchen TestGen is a more focused option when the priority is data testing: its free plan covers one user, one connection, and one project, with unlimited tables and data volume.
Apache Airflow is a straightforward alternative for teams seeking a free Apache-licensed tool. Apache NiFi is another free Apache-licensed choice, with no commercial plan or usage limits stated on its official pages. Pick Apache Hop if an open-source platform requiring Java 21 for release 2.19.0 better matches your environment.
Metaflow is worth considering for teams comfortable scaling production in their own cloud account or Kubernetes cluster. Prefect has a free Hobby plan with two users, one workspace, five deployments, 500 minutes per month on Prefect Serverless, and seven-day run retention, making its stated limits more suitable for a small trial or project. Dagster is another open-source orchestration framework to consider.
Verdict
Flyte is a strong fit for Python teams running AI workloads that need durable recovery, runtime decisions, and cluster-based deployment. Its free OSS runtime provides a no-cost way to run locally or on a team’s infrastructure, and its serving and ML integrations cover a broad range of tools. The main reason to look elsewhere is cost: Team’s $950 monthly minimum plus usage is difficult to justify for modest workloads, and Enterprise is the route for capacity beyond Team’s defined limits.
The verdict
Flyte combines workflow authoring, recovery, dynamic execution, and serving integrations for AI workloads. Its free open-source runtime supports local and cluster deployment, while the Team and Enterprise plans specify additional limits and support.
Questions people ask
Is Flyte free?
Flyte OSS is listed as free under the Apache 2.0 license. The Team plan is $950 per month plus usage and includes a 30-day free trial.
Where can Flyte run?
Flyte 2 can run locally with a devbox or on a user's cluster. The deployment model is listed as self-hosted.
What are the Team plan limits?
Team lists a 25,000 action-per-run fanout limit, 1,000 concurrent actions, one cluster, and 30-day data retention.
What does Enterprise include?
Enterprise lists custom concurrent actions, custom retention, 3+ clusters, and white-glove support. Its pricing is custom.
Which tools integrate with Flyte?
Listed integrations include Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake, Weights & Biases, OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, MLflow, Dask, and Databricks.
What does Union.ai say about data handling?
Union.ai says its data plane runs inside the customer's cloud and that workflow inputs, outputs, logs, code, and reports do not pass through its control plane.
What the maker publishes
- Autoscaling
- Long-running apps can scale replicas up or down automatically and scale to zero when idle. source
- Batch inference
- Flyte 2 includes a DynamicBatcher for batching inference requests with configurable batch size, queue size and timeout. source
- Caching and reproducibility
- Executions are automatically logged and versioned, and identical inputs can reuse cached results. source
- Certifications
- Union.ai's security page lists SOC 2 Type II, SOC 2 Type I, and HIPAA. source
- CLI
- Flytectl is a Go-based cross-platform command-line client for FlyteAdmin, with Homebrew and shell-script installation instructions. source
- Company facts
- LinkedIn lists Union.ai as privately held, founded in 2021, headquartered in Seattle and employing 11–50 people. source
- Company history
- Union.ai's company page says Flyte was open-sourced in 2020 and Union.ai was founded with $10 million in seed funding in 2021. source
- Deployment
- Flyte 2 can run locally with a devbox or on a user's cluster; the site provides `pip install flyte` and `flyte start devbox` setup commands. source
- Deployment architecture
- Legacy deployment documentation describes sandbox, single-cluster and multi-cluster Kubernetes deployment paths, with production use requiring a real object store. source
- Developer requirements
- Legacy development documentation supports Python 3.8 or newer and installs Flytekit with pip. source
- Documentation depth
- The current documentation includes version selectors for Flyte 2 and Flyte 1 plus guides covering tasks, apps, agents, integrations and CI/CD deployment. source
- Durable execution
- Flyte 2 can recover from infrastructure failures including OOM kills, preempted nodes and disappearing GPUs. source
- Dynamic execution
- Flyte supports branching, loops, and runtime decisions for dynamic workflows and agents. source
- Dynamic workflows
- Workflows use ordinary Python control flow for branching, loops and runtime decisions rather than a separate DSL. source
- Extensibility
- Flyte integrations are implemented through a plugin system and the project invites users to build and submit their own integrations. source
- Failure recovery
- Flyte can resume from a failed task and automatically recover from infrastructure failures such as OOM kills, preempted nodes, or disappearing GPUs. source
- Independent praise: Kubernetes fit
- Coding Protocols praises Flyte's typed pipelines, Kubernetes execution and per-task resource control, recommending it for teams already deep on Kubernetes. source
- Independent praise: typing and reproducibility
- Reddit users repeatedly praise Flyte's enforced typing, caching, versioning and reproducibility for machine-learning workflows. source
- Integrations
- The Flyte site lists integrations including Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake, Weights & Biases, OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, MLflow, Dask, and Databricks. source
- License
- The Flyte repositories and SDKs are published under the Apache License 2.0. source
- Limit: ecosystem gaps
- A Reddit reviewer says some features were missing in Flyte 1 and that they still had to store artifacts in MLflow. source
- Limit: Kubernetes dependency
- Independent comparison guidance positions Flyte primarily for teams already operating Kubernetes, implying a poor fit where that platform expertise is absent. source
- Limit: operational complexity
- Reddit reviewers report setup headaches involving permissions, configuration, firewalls and cloud deployment. source
- Logical recovery
- Failed tasks can be retried and resumed with the same inputs instead of rerunning the whole workflow. source
Checked against the maker's pages on 15 Oct 2026.