MosaicML
MosaicML, now integrated with Databricks, is a comprehensive generative AI platform designed for training and deploying large language mode...
Last verified:
What is MosaicML?
Databricks Mosaic AI is a comprehensive platform for developing, training, fine-tuning, and deploying large AI models with high efficiency and scalability. It enables teams to build custom generative AI applications using open-source models like DBRX, MPT, and Shutterstock ImageAI, leveraging advanced tools such as Composer and LLM Foundry for optimized training. The platform supports multi-cloud environments, seamless integration with existing workflows, and focuses on rigorous research to deliver real-world impact in AI innovation.
Key features include DBRX, a state-of-the-art open-source LLM with a sparse mixture-of-experts architecture for superior performance at 36B active parameters; MPT family of models prioritizing quality and efficiency; and Mosaic BERT for cost-effective pretraining at just $20. It offers Mosaic Diffusion for text-to-image generation, StreamingDataset for efficient data loading, and the Evaluation Gauntlet for robust model assessment. Performance optimizations like FP8 on H100 and fast LLM inference make it ideal for large-scale operations.
Designed for enterprises, AI researchers, and developers aiming to become data + AI companies, Mosaic AI is perfect for those needing secure, scalable training on proprietary data without vendor lock-in. It caters to teams building production-grade generative AI, from startups to large organizations like Shutterstock, emphasizing accessibility, speed, and commercial usability of models.
The platform democratizes advanced AI by providing open-source codebases, GitHub repositories, and free trials, allowing users to download models from Hugging Face or train custom versions via Multi-Cloud Training.
MosaicML pricing
Pricing model: Free
Pricing details not explicitly listed on the website; mentions Mosaic BERT pretraining at $20, free trials available via 'Try it free', open-source models downloadable at no cost from Hugging Face, but full platform and Multi-Cloud Training likely usage-based enterprise plans with paid compute.
MosaicML pros
- Trains LLMs with single command simplicity
- DBRX top open-source model quality
- Sparse MoE for fast 36B performance
- MPT-30B prioritizes model quality
- MPT-7B optimized for efficiency
- Mosaic BERT pretrain for $20
- Multi-cloud agnostic deployment
- Composer library for scalability
- LLM Foundry for fine-tuning
- StreamingDataset boosts data efficiency
- Evaluation Gauntlet for quality checks
- FP8 support on H100 GPUs
- Fast LLM inference capabilities
- Shutterstock ImageAI integration
- Open-source commercially usable models
- Handles node failures automatically
- Interoperable with workflows
MosaicML cons
- Proprietary platform requires Databricks
- No visible free tier details
- Enterprise-focused high costs likely
- Dependent on GPU infrastructure
- Limited to supported model families
- Multi-cloud but setup complexity
- Research-heavy less beginner-friendly
- Custom training needs expertise
- Evaluation limited to gauntlet metrics
- ImageAI tied to Shutterstock data
- No on-prem simple option shown
Frequently asked questions about MosaicML
What is DBRX?
DBRX is an open-source, commercially usable LLM developed by Databricks Mosaic AI team, released in March 2024. It features a sparse mixture-of-expert architecture, delivering highest-quality open-source performance with just 36B active parameters. Download it on Hugging Face and use with Databricks Model Serving.
What are MPT models?
MPT models are a family of open-source, commercially usable LLMs released in summer 2023, including MPT-30B for quality prioritization and MPT-7B for efficiency. Users can download pre-trained versions or train custom MPT on their data using Mosaic AI Multi-Cloud Training.
How much does Mosaic BERT cost?
Pretrain your own BERT model from scratch on your data using Mosaic AI for $20. Code is available on GitHub for easy implementation.
What is Composer?
Composer is an open-source deep-learning training library optimized for scalability and usability. It helps compose efficiency methods to speed up training and improve quality. Code available on GitHub.
What is LLM Foundry?
Databricks LLM Foundry is a highly efficient, open-source codebase for training, fine-tuning, and evaluating LLMs. It includes throughput tables for performance benchmarking. Code on GitHub.
What is Shutterstock ImageAI?
Shutterstock ImageAI is a text-to-image diffusion model co-developed by Shutterstock and Databricks, trained exclusively on Shutterstock’s proprietary image repository for photorealistic, high-resolution images from trusted data.
What is Mosaic Diffusion?
Mosaic Diffusion is a generative model that turns text descriptions into images, designed for high efficiency. Full code available on GitHub with a dedicated blog post.
What performance features does Mosaic offer?
Features include FP8 on H100 for training, fast LLM inference, and FP8 for serving. The deep learning stack is optimized for training, fine-tuning, and deploying large models at scale.
What is StreamingDataset?
StreamingDataset is an open-source PyTorch DataLoader that streams training datasets efficiently without loading everything into memory. Download from GitHub with blog post details.
What is the Evaluation Gauntlet?
The Evaluation Gauntlet is a library for evaluating generative language model quality across various benchmarks. Includes README and code on GitHub with explanatory blog post.