Databricks MosaicML is the generative AI division of Databricks, an American data and artificial intelligence software company headquartered in San Francisco, California. It originated from MosaicML, a startup founded in 2021 that developed a platform for training, fine-tuning, and deploying large language models (LLMs) and other deep learning models. Databricks acquired MosaicML in June 2023 for $1.4 billion, integrating its technology into the Databricks Data Intelligence Platform to provide customers with tools for building custom generative AI models on their own data.
The MosaicML platform is known for its emphasis on efficiency and cost reduction in training large neural networks. It includes the MosaicML Composer library, which implements techniques such as curriculum learning, gradient clipping, and layer normalization to accelerate training, and the MosaicML LLM Foundry, a suite of tools for training, fine-tuning, and serving LLMs. After the acquisition, Databricks rebranded these offerings under the Databricks MosaicML name, positioning them as a core component of its managed AI infrastructure.
History
MosaicML was founded in 2021 by Naveen Rao, Hanlin Tang, and Jonathan Frankle. Rao previously co-founded Nervana Systems, a deep learning hardware startup acquired by Intel in 2016. The company aimed to democratize large-scale model training by making it more accessible and affordable for enterprises. In October 2021, MosaicML raised $37 million in a Series A funding round led by Addition, with participation from Future Ventures and others.
In 2022, MosaicML released the MosaicML Composer library, which introduced a set of training methods that could reduce the cost of training deep learning models by up to 50%. The company also launched the MosaicML Cloud, a managed service for training and deploying models on cloud infrastructure. By early 2023, MosaicML had gained attention for its open-source models, including MPT-7B, a 7-billion-parameter transformer-based LLM that was released in May 2023. MPT-7B was notable for its competitive performance against larger models like LLaMA-7B, and it was trained on a dataset of 1 trillion tokens.
In June 2023, Databricks announced its intention to acquire MosaicML for $1.4 billion. The acquisition was completed in July 2023, and MosaicML became a wholly owned subsidiary of Databricks. The deal was part of Databricks' strategy to integrate generative AI capabilities into its lakehouse architecture, which combines elements of data warehouses and data lakes. The acquisition also brought MosaicML's team of researchers and engineers into Databricks, including its CEO Naveen Rao, who became a vice president at Databricks.
Following the acquisition, Databricks continued to develop and release models under the MosaicML brand. In July 2023, the company released MPT-30B, a 30-billion-parameter model, and later that year, it introduced the MPT-7B-Instruct and MPT-30B-Instruct variants fine-tuned for instruction following. In November 2023, Databricks released DBRX, a 132-billion-parameter mixture-of-experts (MoE) model, which was trained using the MosaicML platform. DBRX was designed to compete with models from OpenAI, Anthropic, and Google DeepMind, and it demonstrated strong performance on benchmarks such as MMLU and HellaSwag.
Technology and Platform
The Databricks MosaicML platform is built around the concept of enabling enterprises to train and deploy LLMs on their own data, addressing concerns about data privacy and security. It provides a suite of tools that integrate with Databricks' lakehouse architecture, allowing customers to use their existing data stored in Delta Lake, an open-source project that adds ACID transaction support to data lakes.
Key components of the platform include:
- MosaicML Composer: An open-source library that implements a variety of training techniques, including curriculum learning, gradient clipping, and layer normalization, to improve training efficiency and model quality.
- MosaicML LLM Foundry: A collection of tools for training, fine-tuning, and serving LLMs, with support for popular architectures such as the transformer and for distributed training across multiple GPUs.
- MosaicML Inference: A service for deploying models in production, offering low-latency inference with support for batching and quantization.
These tools are designed to work with major cloud providers, including Amazon Web Services, Microsoft Azure, and Google Cloud Platform. Databricks MosaicML also supports the use of specialized hardware, such as AWS Trainium chips and Intel processors, to reduce training costs.
Models and Research
Databricks MosaicML has released several open-source models that have contributed to the field of generative AI. The MPT (Mosaic Pretrained Transformer) series includes models ranging from 7 billion to 30 billion parameters, all based on the transformer architecture. These models were trained on large text corpora and have been used for tasks such as text generation, summarization, and question answering.
In 2024, Databricks MosaicML introduced DBRX, a 132-billion-parameter MoE model that uses a sparse activation pattern to achieve high performance with lower computational cost. DBRX was trained using the MosaicML platform and was released under an open-source license, allowing researchers and developers to use it freely.
The research team at Databricks MosaicML has also published papers on topics such as efficient training methods, model compression, and the impact of data quality on model performance. They have collaborated with academic institutions, including the University of California, Berkeley, and Carnegie Mellon University, to advance the state of the art in deep learning.
Integration with Databricks Platform
Databricks MosaicML is integrated into the broader Databricks Data Intelligence Platform, which includes the Lakebase database for AI agents, Lakeflow Designer for building data pipelines, and Agent Bricks for developing AI agents. The integration allows customers to use MosaicML's training and inference capabilities directly within their Databricks workspace, with data stored in the lakehouse.
One of the key benefits of this integration is the ability to fine-tune LLMs on proprietary data without moving data to external services. This is particularly important for industries with strict data governance requirements, such as healthcare and finance. Databricks MosaicML also supports the deployment of models for real-time inference, enabling applications like chatbots and virtual assistants.
In 2025, Databricks expanded its generative AI offerings by partnering with OpenAI and Anthropic to incorporate their models into the platform. However, Databricks MosaicML remains a distinct offering for customers who prefer to train and control their own models.
Business and Market Position
Databricks MosaicML competes with other AI infrastructure providers, including Amazon Web Services (with its SageMaker and Trainium offerings), Google Cloud (with Vertex AI), and Microsoft Azure (with Azure Machine Learning). It also competes with startups like AI21 Labs and Inflection AI, which offer proprietary models and APIs.
The acquisition of MosaicML was part of Databricks' broader strategy to become a leading provider of AI infrastructure. Databricks has reported strong revenue growth, with $1.6 billion in revenue for the 2023 fiscal year. The company has also raised significant funding, including a $10 billion round in December 2024 at a $62 billion valuation, and a $4 billion Series L round in December 2025 at a $134 billion valuation.
Impact and Reception
Databricks MosaicML has been well-received by the AI community for its focus on open-source models and cost-efficient training. The release of MPT-7B and DBRX has been praised for making high-quality LLMs accessible to a wider audience. However, some critics have noted that the models may not match the performance of the largest proprietary models from OpenAI or Google DeepMind, and that the platform requires significant technical expertise to use effectively.
Despite these challenges, Databricks MosaicML has established itself as a key player in the enterprise AI market, with a strong emphasis on data privacy and customization. As of 2026, the platform continues to evolve, with ongoing research into new training methods and model architectures.