FLAML
microsoft.github.io
- Free plan
- Windows
- macOS
- Linux
- API
- Self-hosted
What FLAML is
FLAML is a free, MIT-licensed library for automated machine learning and tuning. It searches for accurate models or configurations for common machine-learning and AI tasks while using low computational resources. Its task-oriented AutoML covers classification, regression, time-series and panel forecasting, learning to rank, and sequence classification. The flaml.default package provides data-dependent hyperparameter defaults without runtime tuning. Tuning can account for large search spaces, different evaluation costs, constraints, guidance, and early stopping, while users can limit learners or tune a customized learner. FLAML can adapt large language models to applications to reduce monetary costs. Optional integrations include OpenAI models, CatBoost, Vowpal Wabbit, Prophet, statsmodels, Hugging Face Transformers, Ray, Spark, MLflow, Azure Machine Learning, and Microsoft Fabric. A .NET implementation is available through ML.NET.
Who it suits
FLAML suits developers and data scientists who need automated model selection, tuning, forecasting, ranking, or language-model adaptation. It supports Python, .NET, API, self-hosted, Linux, macOS, and Windows environments.
Good to know, and what to watch
- Free MIT-licensed open-source software
- Supports classification, regression, forecasting, and ranking
- Handles constraints and early stopping during tuning
- Offers Ray and Spark parallel backends
- Provides a .NET implementation through ML.NET
- Latest version requires Python 3.10 to earlier than 3.14
- Ray and Spark cannot run together in one job
- Spark backend does not support GPU training
- Optional integrations require separate installations
The verdict
FLAML combines broad AutoML task coverage with resource-conscious tuning and multiple integration options. Check its Python requirement and Spark GPU limitation before choosing a setup.
What the maker publishes
- Auto-tuning
- Its tuning approach handles large search spaces, heterogeneous evaluation costs, complex constraints, guidance, and early stopping. source
- Availability
- The repository identifies FLAML as MIT-licensed open-source software. source
- Community support
- The project provides community support through Discord, issues, discussions, and contributions. source
- Customization
- Users can restrict the learners FLAML tries or tune a customized learner. source
- Distributed integrations
- FLAML provides optional Ray and Spark distributed tuning, plus NNI, BlendSearch, and Azure Synapse options. source
- Installation
- The Python library can be installed with pip or conda, and optional packages enable features such as AutoML, Hugging Face Transformers, Ray, and Spark. source
- Integrations
- The documentation describes MLflow logging and integration with Azure Machine Learning. source
- License
- The repository is distributed under the MIT License, permitting users to use, copy, modify, publish, distribute, sublicense, and sell the software subject to its conditions. source
- LLM adaptation
- FLAML automatically adapts large language models to applications to reduce monetary costs. source
- Microsoft Fabric
- The repository says FLAML supports AutoML and hyperparameter tuning in Microsoft Fabric Data Science. source
- Model integrations
- Optional installations include OpenAI models, CatBoost, Vowpal Wabbit, Prophet, statsmodels, and Hugging Face Transformers. source
- .NET support
- FLAML has a .NET implementation in the cross-platform ML.NET framework, including Model Builder, the ML.NET CLI, and Microsoft.ML.AutoML. source
- Parallel tuning
- FLAML supports Ray and Spark backends for parallel tuning, but a tuning job cannot use both. source
- Purpose
- FLAML is a fast library for automated machine learning and tuning. source
- Purpose
- FLAML automates machine learning model selection and hyperparameter tuning with low computational resources. source
- Python requirement
- The installation documentation requires Python version 3.10 or newer. source
- Resource efficiency
- It finds accurate models or configurations for common ML/AI tasks with low computational resources. source
- Runtime requirement
- The repository specifies Python 3.10 or later and earlier than 3.14 for its latest version. source
- Security reporting
- The project directs security vulnerability reports to Microsoft Security Response Center for coordinated disclosure rather than public issues. source
- Spark limitation
- GPU training is not supported when Spark is used as the parallel backend. source
- Support
- The official site links to a Discord community. source
- Supported tasks
- Task-oriented AutoML supports classification, regression, time-series forecasting, panel forecasting, learning to rank, and sequence classification. source
- Tasks
- The documentation lists classification, regression, forecasting, and ranking among the supported AutoML tasks. source
- Tuning
- Its tuning tool handles large search spaces with varied evaluation costs, constraints, guidance, and early stopping. source
Checked against the maker's pages on 2 Oct 2026.