Citation and evidence

MosaicML

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI CompaniesLarge Language ModelsOpen Source AI

Cite this article

MosaicML was an American generative AI startup, based in San Francisco, that built software and cloud infrastructure for training large language models and other neural networks at lower cost. Founded in 2021 by Naveen Rao and Hanlin Tang, two former Intel executives, it released the open-source training library Composer, published estimates of the cost of training models such as GPT-3 and Stable Diffusion, and in 2023 released the MPT family of open, commercially usable language models.[1][4][11][13][14] Databricks announced on June 26, 2023 that it would acquire MosaicML for about $1.3 billion, a figure that included retention packages, and completed the purchase on July 19, 2023.[5][6] MosaicML's research and engineering team joined Databricks,[5][6] which later used the Mosaic name for its AI product line, Mosaic AI.[20]

Overview

FieldValue
TypePrivate company (acquired)
Founded2021
FoundersNaveen Rao, Hanlin Tang (co-founders); Jonathan Frankle and Michael Carbin also named as co-founders by MIT News
HeadquartersSan Francisco, California
Public launch2021-10-13 (left stealth with $37 million)
Total fundingJust under $64 million
Last private valuation$222 million
Employees at acquisition62
Notable productsComposer, Streaming, LLM Foundry, MosaicML Training, MosaicML Inference, MPT models
FateAcquired by Databricks (announced 2023-06-26, completed 2023-07-19)

Expanded article table

Sources for the table: founding year and headcount from Fortune;[9] founders from the Kempner Institute and MIT News;[3][4] headquarters from Crunchbase News;[8] launch and first round from SiliconANGLE;[1] funding and valuation from TechCrunch;[7] acquisition dates from Databricks.[5][6]

History

Founding and launch (2020-2021)

Naveen Rao had founded the deep-learning chip company Nervana Systems, which Intel acquired in 2016, and Hanlin Tang had worked with him there; both left Intel the year before MosaicML emerged from stealth.[1][2] According to MIT News, Rao emailed Jonathan Frankle, then a PhD student at the Massachusetts Institute of Technology, and MIT computer scientist Michael Carbin in 2020, after reading a paper the two had written on shrinking deep-learning models, and pitched them on starting a company; Tang joined them, and MIT News counts Frankle and Carbin as co-founders alongside Rao.[3] Harvard's Kempner Institute, where Frankle was later affiliated, describes him as MosaicML's founding adviser and chief scientist and names Tang as chief technology officer.[4]

MosaicML came out of stealth on October 13, 2021 with $37 million from investors including Lux Capital, DCVC, Future Ventures, Playground Global, AME, Correlation and E14, plus several angel investors.[1] MIT's E14 Fund was an investor in the Series A round.[3] The company's pitch was that the cost of training neural networks could be cut through algorithmic and systems-level changes rather than new hardware.[2] At launch it presented two tools: Composer, an open-source library of training methods (about 20 at the time, for computer vision and natural language processing) that could be combined into "recipes", and MosaicML Explorer, an interface for comparing the time, cost and accuracy of training runs across cloud services and hardware.[1][2] MosaicML reported that its methods trained ResNet-50 on ImageNet 2.9 times faster and more cheaply than an optimized baseline at the same accuracy, ResNet-101 3.5 times faster, and a 125-million-parameter GPT-style language model 1.7 times faster, measured on eight NVIDIA A100 GPUs on AWS.[2]

Cost studies and the training platform (2022 to early 2023)

In 2022 and 2023 MosaicML published estimates of what it cost to train large models on its platform. In a September 29, 2022 post, MosaicML engineers Abhi Venigalla and Linden Li argued that a model reaching the quality of the original 175-billion-parameter GPT-3 could be trained for about $450,000, "2x-10x less than people think".[13] The estimate was a projection: it used the Chinchilla scaling laws to propose a 30-billion-parameter model trained on 610 billion tokens, which those laws predicted would reach a slightly lower loss than GPT-3 with about 65% less compute, and the authors noted that the final quality of such a model was "still a projection".[13]

A second line of work concerned image generation. On April 26, 2023, a MosaicML team reported training a model comparable to Stable Diffusion 2 base from scratch in 7.45 days for under $50,000 (23,835 A100 hours at an assumed $2 per hour), using 128 A100 GPUs, Composer, the company's StreamingDataset library and its orchestration platform; they said this was about 8 times cheaper than the original Stable Diffusion 2. They released the training code but not the weights, citing legal uncertainty around models trained on LAION-5B.[14]

Inference service and MPT models (2023)

On May 3, 2023, MosaicML launched MosaicML Inference, a hosted service for deploying curated open-source models for text completion, text embedding and image generation. The company claimed text completion and embeddings could be served at a quarter of the cost of comparable OpenAI models, and image generation at one fifteenth the cost of DALL-E 2.[16]

Two days later, on May 5, 2023, it released MPT-7B, a 6.7-billion-parameter decoder-only transformer trained on 1 trillion tokens of text and code. MosaicML said the model was trained in about 9.5 days on 440 A100-40GB GPUs in Oracle Cloud Infrastructure for about $200,000, with no human intervention (four hardware failures were detected and the run resumed automatically), and that it matched the quality of Meta's LLaMA-7B.[11] The base model carried the Apache 2.0 license, which allows commercial use.[11] The release also included three fine-tuned variants, among them MPT-7B-StoryWriter-65k+, which was tuned on a 65,000-token context and demonstrated generating up to 84,000 tokens using ALiBi position encoding.[11] The architecture combined ALiBi with FlashAttention and used the Lion optimizer instead of AdamW.[11]

MPT-30B followed on June 22, 2023. MosaicML said the 30-billion-parameter model outperformed the original GPT-3 on six of nine in-context learning metrics while using about 17% of GPT-3's parameters, and that it could run on a single A100-80GB GPU at 16-bit precision. Most of its training ran on A100 clusters; after receiving 256 NVIDIA H100 GPUs from CoreWeave in mid-June the company finished the run on H100s, and it described MPT-30B as the first public model to be partly trained on H100s.[12] By that date MosaicML said the MPT-7B models had been downloaded more than 3 million times.[12] The company also showed that its stack could train LLMs on AMD Instinct MI250 GPUs with PyTorch 2.0 and ROCm without code changes, reaching about 80% of the per-GPU throughput of an A100-40GB on a single four-GPU node.[15]

Acquisition by Databricks (2023)

According to Fortune, Databricks chief executive Ali Ghodsi and Rao first met on March 30, 2023 at the Cerebral Valley conference in San Francisco; merger talks began with a video call on May 10, and Ghodsi signed a term sheet on May 29.[10] Databricks announced a definitive agreement on June 26, 2023, valuing the deal at approximately $1.3 billion including retention packages, and said MosaicML's research and engineering team would join Databricks.[5] The announcement cited 3.3 million downloads of MPT-7B and named AI2 (Allen Institute for AI), Generally Intelligent, Hippocratic AI, Replit and Scatter Labs as MosaicML customers.[5]

Coverage at the time stressed the size of the price relative to the company. TechCrunch reported that MosaicML had raised just under $64 million and that its last private valuation was $222 million, about a sixth of the purchase price, and suggested that the retention arrangement for the whole team "likely could be one reason for the large price tag".[7] Fortune described it as a stock deal and reported that all 62 MosaicML employees were expected to stay.[9]

The acquisition closed on July 19, 2023.[6] Databricks said MosaicML's product would remain available for direct purchase and would be integrated with its lakehouse platform, and that it would keep working with MosaicML's research team to innovate "in the open" on model architectures like those behind MPT-7B and MPT-30B.[6][18] The day before the close, MosaicML published MPT-7B-8K, an MPT-7B continuation trained on 500 billion more tokens with an 8,000-token context, in three days on 256 H100 GPUs.[19]

Products and open-source software

MosaicML sold managed training and inference and published most of its tooling under the Apache 2.0 license.

ProductTypeDescription
ComposerOpen-source libraryPyTorch-based training library for distributed training on large clusters, distributed on PyPI as mosaicml; used to train the MPT models[22]
StreamingDataset (Streaming)Open-source libraryData streaming library for neural network training, with deterministic shuffling and mid-epoch resumption[14][24]
LLM FoundryOpen-source codebaseTraining and fine-tuning code used to build MPT-7B and MPT-30B[12][23]
MosaicML ExplorerWeb toolCompared cost, time and accuracy of training runs across hardware and clouds (2021)[1]
MosaicML TrainingManaged serviceFine-tuning, domain-specific pretraining or training from scratch, with the customer owning the weights[12]
MosaicML InferenceManaged serviceHosted endpoints for open models, launched 2023-05-03[12][16]
MCLICommand-line toolOrchestration of runs on the MosaicML platform[11]

Expanded article table

The MPT-7B release described the stack used to train it: Composer, PyTorch FSDP, LLM Foundry, StreamingDataset and the MosaicML platform's MCLI orchestration tool.[11]

Models

ModelReleasedParametersLicense (base model)Notes
MosaicBERT (base and large)2023Base and large sizesApache 2.0 (base checkpoints on Hugging Face)Fast-pretraining BERT encoder recipe; paper at NeurIPS 2023[17][25]
MPT-7B2023-05-056.7BApache 2.01T tokens; Instruct, Chat, StoryWriter-65k+ variants[11]
MPT-30B2023-06-2230BApache 2.08k context; partly trained on H100s; Instruct and Chat variants[12]
MPT-7B-8K2023-07-186.7BCC BY-SA 3.01.5T tokens total, 8k context[19]

Expanded article table

The fine-tuned variants did not all share the base license. MPT-7B-Instruct carried CC BY-SA 3.0, while MPT-7B-Chat and MPT-30B-Chat carried the non-commercial CC BY-NC-SA 4.0 license, and MosaicML described MPT-30B-Chat as a research artifact not meant for commercial use.[11][12] The MPT article covers the model family in more detail.

Research

Besides its blog posts, the company published peer-reviewed research. MosaicBERT, presented at NeurIPS 2023, is a BERT-style encoder recipe that adds FlashAttention, ALiBi, gated linear units, removal of padding tokens and low-precision LayerNorm, together with a 30% masking ratio and bfloat16 training. The authors reported that the base model reached an average GLUE (dev) score of 79.6 after 1.13 hours of pretraining on eight A100 80 GB GPUs, at a cost of roughly $20, and released weights and code.[17] As of November 2026 the five MosaicBERT checkpoints are the only models listed on the "Mosaic ML, Inc." organization page on Hugging Face; the MPT repositories no longer appear there.[25]

Customers also used the platform to train their own models. Replit's replit-code-v1-3b, a 2.7-billion-parameter code-completion model published on Hugging Face in April 2023, was trained on the MosaicML platform on 256 A100-40GB GPUs using the company's examples repository.[26]

Legacy

Inside Databricks the former MosaicML group led the development of DBRX, a mixture-of-experts model with 132 billion total and 36 billion active parameters released on March 27, 2024. According to Databricks, DBRX's development "was led by the Mosaic team that previously built the MPT model family". The model was trained with the company's open-source libraries MegaBlocks, LLM Foundry, Composer and Streaming, and Databricks said its training recipe could match the quality of the MPT models with nearly four times less compute.[27] Frankle told MIT News that DBRX mattered because it was built with Databricks tools, so the company's customers could reach similar performance with their own models.[3]

Databricks also adopted the name for its AI products. By its Data + AI Summit in June 2024, TechCrunch described the platform as "rebranded as Mosaic AI", with new tools including an agent framework, agent evaluation, a tools catalog, model training and an AI gateway; an analyst quoted by InfoWorld observed that Databricks had begun branding most of its generative AI capabilities with the Mosaic name.[20][21] TechCrunch reported at the time of the deal that Rao would stay on with the team.[7] He spent a little over two years as Databricks' vice president of AI and left in September 2025 to start a hardware company, Unconventional, Inc.[28]

As of November 2026, GitHub marks the Composer and LLM Foundry repositories as archived (read-only), while the Streaming repository has not been archived.[22][23][24]

Continue reading

More from See also

More to read

Topics:AI CompaniesLarge Language ModelsOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6Wheatley, Mike. "Naveen Rao's MosaicML debuts with a mission to improve machine learning training". SiliconANGLE, 2021-10-13. siliconangle.com/...th-mission-improve-ml-training
  2. ^1 ^2 ^3 ^4Prickett, Nicole Hemsoth. "Former Nervana Leads Target Optimal Training Configurations". The Next Platform, 2021-10-13. nextplatform.com/...1644816
  3. ^1 ^2 ^3 ^4Winn, Zach. "Helping nonexperts build advanced generative AI models". MIT News, 2024-06-21. news.mit.edu/...advanced-generative-ai-models-0621
  4. ^1 ^2 ^3Kempner Institute, Harvard University. "MosaicML Acquisition: Generative AI Platform Developed by Kempner's Jonathan Frankle Acquired by Databricks in $1.3 Billion Deal". Kempner Institute, 2023-06-30. kempnerinstitute.harvard.edu/...saicml-acquisition
  5. ^1 ^2 ^3 ^4 ^5Databricks. "Databricks Signs Definitive Agreement to Acquire MosaicML, a Leading Generative AI Platform". Databricks Newsroom, 2023-06-26. databricks.com/...l-leading-generative-ai-platform
  6. ^1 ^2 ^3 ^4 ^5Databricks. "Databricks Completes Acquisition of MosaicML". Databricks Newsroom, 2023-07-19. databricks.com/...s-completes-acquisition-mosaicml
  7. ^1 ^2 ^3Lunden, Ingrid. "Databricks picks up MosaicML, an OpenAI competitor, for $1.3B". TechCrunch, 2023-06-26. techcrunch.com/...ml-an-openai-competitor-for-1-3b
  8. ^Metinko, Chris. "Databricks Buys Generative AI Startup MosaicML For $1.3B". Crunchbase News, 2023-06-26. news.crunchbase.com/...startup-databricks-mosaicml
  9. ^1 ^2Beltran, Luisa. "Databricks says it's buying MosaicML in a bid to 'democratize A.I.'". Fortune, 2023-06-27. fortune.com/...osaicml-in-a-bid-to-democratize-a-i
  10. ^Beltran, Luisa. "The inside story of how one of the biggest A.I. deals of the year came together, told by the CEOs". Fortune, 2023-07-21. fortune.com/...-biggest-ai-deal-came-together-ceos
  11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9The Databricks AI Research Team (originally MosaicML). "Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs". Databricks Blog, 2023-05-05. databricks.com/...mpt-7b
  12. ^1 ^2 ^3 ^4 ^5 ^6 ^7The Databricks AI Research Team (originally MosaicML). "MPT-30B: Raising the bar for open-source foundation models". Databricks Blog, 2023-06-22. databricks.com/...mpt-30b
  13. ^1 ^2 ^3Venigalla, Abhi; Li, Linden. "Mosaic LLMs: GPT-3 quality for <$500k". Databricks Blog (originally MosaicML Blog), 2022-09-29. databricks.com/...gpt-3-quality-for-500k
  14. ^1 ^2 ^3Patel, Mihir; Stephenson, Cory; Seguin, Landan; Jacobson, Austin; Yuen, Erica Ji. "Training Stable Diffusion from Scratch for <$50k with MosaicML (Part 2)". Databricks Blog (originally MosaicML Blog), 2023-04-26. databricks.com/...stable-diffusion-2
  15. ^Venigalla, Abhi. "Training LLMs with AMD MI250 GPUs and MosaicML". Databricks Blog (originally MosaicML Blog), 2023-06-30. databricks.com/...amd-mi250
  16. ^1 ^2Dotson, Kyt. "MosaicML launches Inference to reduce enterprise generative AI deployment costs". SiliconANGLE, 2023-05-03. siliconangle.com/...generative-ai-deployment-costs
  17. ^1 ^2Portes, Jacob; Trott, Alex; Havens, Sam; King, Daniel; Venigalla, Abhinav; Nadeem, Moin; Sardana, Nikhil; Khudia, Daya; Frankle, Jonathan. "MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining". NeurIPS 2023; arXiv:2312.17482, 2023-12-29. arxiv.org/...2312.17482
  18. ^Zaharia, Matei; Wendell, Patrick; Xin, Reynold; Ghodsi, Ali. "Databricks + MosaicML". Databricks Blog, 2023-07-19. databricks.com/...databricks-mosaicml
  19. ^1 ^2Havens, Sam; Yuen, Erica Ji. "Announcing MPT-7B-8K: 8K Context Length for Document Understanding". Databricks Blog, 2023-07-18. databricks.com/...long-context-mpt-7b-8k
  20. ^1 ^2Lardinois, Frederic. "Databricks expands Mosaic AI to help enterprises build with LLMs". TechCrunch, 2024-06-12. techcrunch.com/...help-enterprises-build-with-llms
  21. ^Ghoshal, Anirban. "Databricks expands Mosaic AI support for generative AI apps". InfoWorld, 2024-06-12. infoworld.com/...ai-support-for-generative-ai-apps
  22. ^1 ^2MosaicML. "mosaicml/composer: Supercharge Your Model Training". GitHub, accessed 2026-11-18. github.com/...composer
  23. ^1 ^2MosaicML. "mosaicml/llm-foundry: LLM training code for Databricks foundation models". GitHub, accessed 2026-11-18. github.com/...llm-foundry
  24. ^1 ^2MosaicML. "mosaicml/streaming: A Data Streaming Library for Efficient Neural Network Training". GitHub, accessed 2026-11-18. github.com/...streaming
  25. ^1 ^2Hugging Face. "Mosaic ML, Inc." (mosaicml organization page). Hugging Face, accessed 2026-11-18. huggingface.co/mosaicml
  26. ^Replit. "replit-code-v1-3b" (model card). Hugging Face, 2023. huggingface.co/...replit-code-v1-3b
  27. ^The Databricks AI Research Team. "Introducing DBRX: A New State-of-the-Art Open LLM". Databricks Blog, 2024-03-27. databricks.com/...cing-dbrx-new-state-art-open-llm
  28. ^Temkin, Marina. "Sources: Naveen Rao's new AI hardware startup targets $5B valuation with backing from a16z". TechCrunch, 2025-10-03. techcrunch.com/...valuation-with-backing-from-a16z

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 2,581 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "MosaicML." aiwiki.ai, updated 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/mosaicml

Edit article