Bria Trains AI Models Responsibly Using Amazon SageMaker AI

Bria, an AWS Partner, is a software startup that launched in 2020 and provides enterprises with the building blocks they need to build and enhance custom generative AI solutions. Through a full suite of models, auxiliary models, control nets, adapters, and algorithms, Bria empowers its customers to solve real-world business challenges using generative AI. Moreover, the company grants its customers full access to source code, weights, and comprehensive documentation for its trained models.

The company is fully committed to using AI responsibly. It trains its FMs from scratch using data that is specifically licensed for generative AI purposes. Also, Bria uses proprietary and patented attribution technology to compensate content creators on the basis of their contribution to the generated content. Bria’s models are trained only on licensed and commercial-grade data and do not include trademarks, logos, harmful content, or misinformation.

Additionally, Bria proactively pursues diverse, comprehensive content to reduce the potential for bias formation. “Since its inception, Bria has been pioneering the responsible AI space,” says Vered Horesh, chief of strategic AI partnerships at Bria. “We are deeply involved in what it means to be an accountable AI developer, and we are constantly evolving our commitment to this philosophy.” The company’s ethical approach effectively removes the core privacy, legal, copyright, and accuracy concerns that enterprises have about adopting generative AI technology.

Training its models using in-house infrastructure would have demanded significant time and personnel—a significant burden on startup resources. So Bria decided to use a managed solution to help its team focus on training models without having to handle ML operations. The company chose Amazon SageMaker AI as a flexible and scalable solution. “We don’t constantly occupy compute for the way we train our models; there are a lot of pauses,” says Bar Fingerman, head of AI and ML engineering at Bria. “Using Amazon SageMaker is very cost-efficient for us. When we’re not training, we’re not paying for the compute. We can also allocate jobs very simply on demand.”

Bria built its initial model-training stack in 2020 using Amazon SageMaker AI model training capabilities. The process involves three steps: First, the company gathers raw licensed training data and stores it using Amazon Simple Storage Service (Amazon S3), an object storage service built to retrieve virtually any amount of data from anywhere. Second, it uses several other AWS services to onboard, process, analyze, and structure this growing volume of data. Finally, Bria trains its models on Amazon SageMaker AI using distributed data parallel policy. The company powers the models with Amazon Elastic Compute Cloud (Amazon EC2) P4d Instances, specifically P4de Instances, which deliver high performance for ML training in the cloud. Bria takes 1–2 months to train an FM, though it might take 2 weeks to train smaller models.

Bria has trained more than 50 FMs. The company’s models exist in AWS Marketplace, where organizations can find, buy, and immediately start using the software and services that run on AWS. The models are also available on Amazon SageMaker JumpStart, an ML hub with FMs, built-in algorithms, and prebuilt ML solutions that users can deploy with a few clicks. Customers use their own proprietary data for additional training. For example, a top gaming company can use game data to tweak models for use in 3D pipelines. Therefore, compliance and security are top priorities for Bria and its customers. Using Amazon SageMaker, the company can help keep intellectual property risk-free, private, and secure under the AWS Shared Responsibility Model and Bria’s full intellectual-property indemnity. “Having the fully managed and secured environment of Amazon SageMaker AI is critical for our customers,” says Horesh.

Furthermore, Bria’s models use Amazon SageMaker AI to lower the entry barrier for end users. Most users are researchers without technical training, and Bria’s models help them perform complex tasks with the click of a button. For customers, this means a significant reduction in time to market. “When customers onboard Amazon SageMaker AI, they can simply plug and play,” says Fingerman. “Customers can get our solutions and instantly start engaging instead of spending time setting up new models.”

As the market moves toward responsible generative AI solutions, Bria has seen consistent year-over-year growth. In 2023–2024, the company’s customer base increased by five times. Bria can seamlessly scale to meet this growing demand. “Using Amazon SageMaker AI, we can support virtually any amount of compute,” says Fingerman. “We started with 4 machines, and today we’re up to 32. We’re ready to scale even more; it’s effortless.”

Using Amazon SageMaker AI, Bria built FMs to meet the rising demand for responsible generative AI solutions. As more enterprises emphasize the importance of responsible AI, Bria plans to continue working alongside AWS to develop new solutions and drive the industry forward.

“We largely focus on helping businesses actually take our models into production,” says Horesh. “With the advanced security features, reduced time to market, and managed environment we get on Amazon SageMaker AI, we can offer our customers the building blocks they need to address real-world use cases.”