Weights & Biases is the AI developer platform for training and fine-tuning models: experiment tracking, model and data versioning, hyperparameter sweeps and production model management in one workspace. This repository is its open-source client library, used by ML teams everywhere, with about 11,300 stars on GitHub.
Serious model development produces chaos โ hundreds of runs, dozens of dataset versions, hyperparameters nobody wrote down. Weights & Biases imposes order with a few lines of code: log metrics, configs, media and system stats during training and every run lands in a shared, searchable dashboard where teams compare results, discuss them and promote the best models to a registry. Artifacts version the datasets and models flowing through a pipeline with full lineage, Sweeps automate hyperparameter search across machines, and the platform extends to LLM workflows with tracing and evaluation tools for prompts and fine-tuning jobs. The client integrates with PyTorch, TensorFlow, Keras, JAX, Hugging Face and most training stacks, runs alongside jobs on any infrastructure, and can be self-hosted for regulated environments. It has become default infrastructure at research labs and AI companies alike, and the client repository remains in daily active development through 2026.
The wandb client library is open source under the MIT license, and the hosted platform has a free tier for individual researchers. Teams and enterprises pay for collaboration, private hosting and scale, so the overall offering is freemium.
Weights & Biases is for ML researchers and engineering teams who train models seriously and need reproducibility, comparability and collaboration across many experiments โ from academic labs to production AI organizations.