Weights & Biases passes one billion tracked runs, CoreWeave says

From a karate studio to a billion runs

Weights & Biases was founded in 2017 with the aim of building the best tools for machine learning practitioners. According to the company, its first product was an experiment tracker built in a karate studio in San Francisco. Teams such as OpenAI, Toyota Research and Uber picked it up quickly, replacing TensorBoard or internal solutions that the company describes as not quite up to the task.

The post, which carries the CoreWeave brand after the two companies joined, thanks users ranging from students building a first computer vision model on MNIST to researchers working on the latest foundation models. The company says its focus over the past nine years has not changed.

Preparing for the road to a trillion runs

CoreWeave says a single foundation-scale effort can now log hundreds of thousands of experiments in a year, a scale it calls impossible in 2017. It expects the experiment load to grow massively as labs use AI in their own training.

As part of that preparation, the company recently launched CoreWeave ARIA, short for AI Research and Iteration Agent. According to the post, ARIA lives in Weights & Biases and helps analyze experiments, generate hypotheses, launch new runs and evaluate results, with the aim of automating the AI research loop.

The company also argues that the people involved in model development have changed. Engineers now ship agents, domain experts review outputs and internal teams build and tweak agents, so far more people take part in the process than in 2017. CoreWeave says the tools and the infrastructure now work as a single system, covering training and inference, tracking and evaluation, and observability and feedback, down to the hardware.

Customer examples behind the number

The post highlights four customers. All descriptions below are the company's own.

• Decart developed Oasis 3, described as the first API accessible world model for physical AI, using W&B Models for experiment tracking, hyperparameter optimization, versioning and lineage tracking.
• Pinterest standardized machine learning development around Weights & Biases, integrated into its MLEnv framework. Hundreds of ML engineers track experiments across hundreds of thousands of training jobs each month.
• Scaled Cognition trains LLMs for regulated industries such as banking and healthcare. The post claims its models achieve 114% higher accuracy than leading general purpose LLMs for enterprise customer support.
• MasterClass uses W&B Weave to evaluate and improve AI teaching agents, tracing full conversations and monitoring production behavior.

What it means for AI teams

The milestone reflects how much experiment tracking has grown alongside AI development. CoreWeave positions Weights & Biases as a tool for both researchers and the wider group of engineers and domain experts who now maintain and evaluate models and agents. With ARIA, it is betting that more of the research loop will be automated and that run counts will keep climbing.

The company says it will keep building tools for AI teams of every size. It also invites users to Fully Connected 2026, held from September 29 to October 1 at Moscone South in San Francisco, to see the latest Weights & Biases work on model and agent development.