Weights & Biases, Inc. is an American company specializing in software for teams developing artificial intelligence models and applications. Based in San Francisco and accessible at wandb.ai, it helps researchers and engineers document their experiments, compare their results and organize their collaborative work. Its focus falls within MLOps, which structures the lifecycle of machine learning models, and extends to generative AI applications.
From laboratory experiments to the CoreWeave ecosystem
Founded in 2017 by Lukas Biewald, Chris Van Pelt and Shawn Lewis, Weights & Biases was built around a practical problem: keeping a usable record of the many trials needed to train a model. When datasets, parameters and code versions change simultaneously, identifying the source of an improvement becomes difficult. The company developed tools to centralize this information and make it easier to access.
In 2025, its acquisition by CoreWeave marks a new stage. The deal brings together a software environment used by AI developers and a provider of GPU computing capacity. It places Weights & Biases within a broader group covering more stages of the development chain.
Tracking, comparing and evaluating models
The platform makes it possible to record training metrics, configurations and files produced during an experiment. Its dashboards make it easier to compare trials, while its sharing and reporting features allow a team to access the results. Integrations with machine learning frameworks connect this tracking to the code developers already use.
Weights & Biases also offers features for hyperparameter optimization, artifact management and a model registry. The goal is to preserve the links between data, versions and results to better understand how a model was produced. This traceability facilitates reproducibility and collaboration, without guaranteeing the quality of the final system on its own.
With Weave, the company extends this approach to applications based on large language models. The tool makes it possible to trace calls, examine interactions between components and conduct evaluations. It addresses a need distinct from simply measuring training performance: checking how a generative application behaves in real-world situations.
What comes next?
The challenge for Weights & Biases now is to connect experimentation, evaluation and the operation of AI applications more closely. Its integration into CoreWeave may make it easier to move between software tools and computing resources. It also raises the question of maintaining an experience suited to organizations using multiple infrastructure providers. In a market where cloud platforms offer their own tools, the ability to remain interoperable, document results and make evaluations useful to teams will be an important differentiator.