tools
Ludwig
Open-source framework for training, fine-tuning, evaluating, and serving custom AI models with YAML configuration.
Ludwig is a declarative deep learning framework for building models across text, tabular data, images, audio, time series, and other modalities. Users define pipelines in YAML, then train, evaluate, export, and serve models through its CLI or Python API.
It is used by machine-learning engineers, researchers, and data scientists for prototyping, multimodal modeling, LLM fine-tuning, hyperparameter optimization, distributed training, and production serving. Ludwig itself is free and open source, but compute, cloud infrastructure, external model providers, and hosted services may cost extra.
Features
- Declarative YAML configuration for preprocessing, modeling, training, and evaluation
- Multimodal and multi-task learning across text, tabular, image, audio, and time-series data
- LLM fine-tuning with SFT, DPO, KTO, ORPO, GRPO, LoRA, and QLoRA
- Distributed training with Ray, DDP, FSDP, DeepSpeed, and KubeRay
- Built-in hyperparameter optimization with Ray Tune and Optuna
- REST and OpenAI-compatible serving for trained models
- Model export to SafeTensors, ONNX, torch.export, and Hugging Face Hub
- Open source under the Apache 2.0 License
Use cases
- Train text classifiers, regressors, summarizers, translators, and question-answering models
- Fine-tune LLMs and vision-language models on custom datasets
- Combine text, images, audio, and tabular data in one model
- Run hyperparameter searches and compare model experiments
- Scale training from a local machine to Ray or Kubernetes clusters
- Serve trained models through REST or OpenAI-compatible endpoints
Pros
Cons
Latest updates
- v0.17.10 (v0.17.10)
Ray workers now return weights as SafeTensors and other results as JSON; the driver no longer unpickles checkpoint data.
- v0.17.9 (v0.17.9)
Checkpoint loading now uses PyTorch’s restricted unpickler to prevent code execution from crafted checkpoints.
- v0.17.8: path traversal fix in dataset archive extraction (v0.17.8)
Fixed path traversal in dataset archive extraction through symlink members.
- v0.17.7 (v0.17.7)
Fixed Ray preprocessing tests for Arrow-backed data and row ordering.
- v0.17.6 (v0.17.6)
Added preprocessing progress callbacks for pandas, Dask, and Ray backends.
Capabilities
- Command line — “$ ludwig serve --model_path results/” source
- Choice of models — “Run Llama, Mistral, Qwen, or any HuggingFace model with 4-bit QLoRA on a single consumer GPU.” source
- Self-hosted — “Train locally with CPU or GPU. Fast iteration, no setup.” source
- API — “One command to serve your model as a REST API.” source
- Official SDKs — “train.py — full Python API” source
Get it
Pricing
- Prices checked
- 2026-11-05