MosaicML was an artificial intelligence company focused on developing software for training and deploying large language models and other machine learning models. The company was known for its open-source libraries and its emphasis on making deep learning more accessible and cost-effective for enterprises. Its technology aimed to reduce the computational resources required for training neural networks, enabling organizations to build custom models with their own data.
Founded in 2021, MosaicML was headquartered in San Francisco, California. The company's core offerings included the MosaicML Composer library for model training and the MosaicML Inference service for deployment. These tools supported a range of model architectures, including transformers, and were designed to work across various cloud platforms. In June 2023, MosaicML was acquired by Databricks, a data and AI company, for approximately $1.4 billion, marking a significant step in Databricks' expansion into generative AI.
Founding and Early Development
MosaicML was established by Naveen Rao and Hanlin Tang, who had previously worked on AI hardware and software. Rao, a former Intel executive, and Tang, a former AI researcher, aimed to address the high costs and complexity associated with training large-scale models. The company quickly gained attention for its innovative approaches to optimizing training processes, such as efficient data loading and model architecture tweaks.
In its early years, MosaicML released several open-source tools that gained traction in the machine learning community. The Composer library, introduced in 2022, provided a collection of training methods and techniques that could accelerate model training by up to 7 times compared to standard approaches. This was achieved through methods like layer normalization, gradient clipping, and curriculum learning, which were integrated into a simple API.
Technology and Products
MosaicML's technology centered on making deep learning more efficient. The company developed a training platform that could be deployed on-premises or in the cloud, supporting major providers like Amazon Web Services, Microsoft Azure, and Google Cloud. Its key product, MosaicML Composer, offered a flexible framework for training neural networks with built-in optimizations.
The company also provided MosaicML Inference, a service for deploying trained models at scale. This service supported various model types, including large language models, and offered features like low-latency responses and cost management. MosaicML's approach allowed enterprises to fine-tune models on their proprietary data, addressing privacy and customization needs that were not easily met by general-purpose models from providers like OpenAI or Anthropic.
Acquisition by Databricks
In June 2023, Databricks announced its intent to acquire MosaicML for $1.4 billion. The acquisition was completed later that year, integrating MosaicML's team and technology into Databricks' platform. This move was part of Databricks' strategy to enhance its generative AI capabilities, allowing customers to build and deploy custom large language models within the Databricks environment.
The acquisition was seen as a response to the growing demand for enterprise AI solutions. By incorporating MosaicML's tools, Databricks aimed to offer a more comprehensive data intelligence platform, combining its lakehouse architecture with advanced model training and deployment capabilities. The integration enabled Databricks to support open-source models and provide alternatives to proprietary systems, giving customers more control over their AI initiatives.
Impact and Legacy
MosaicML's influence extended beyond its acquisition. Its open-source contributions, particularly the Composer library, continued to be used by developers and researchers in the machine learning community. The company's focus on efficiency helped popularize techniques for reducing training costs, which became increasingly important as models grew in size and complexity.
Following the acquisition, MosaicML's technology became a foundational component of Databricks' AI offerings. The tools were integrated into Databricks' platform, enabling users to train and deploy models using their own data within a secure and managed environment. This legacy contributed to Databricks' growth, which saw its valuation rise to $62 billion in December 2024 and further to $134 billion by December 2025, driven in part by the demand for generative AI solutions.