How KREA Scales AI Model Training with WEKA

KREA replaced Ceph with WEKA NeuralMesh™ and saw up to 50x throughput, zero training failures, and faster model releases: https://weka.io/customers/krea-ai/ KREA is an AI creative platform for visual professionals, building state of the art foundation models in house. Before WEKA, their storage broke constantly and researchers lost hours at a time to outages and data loss scares. In this customer story, the KREA team explains how a 30 day proof of concept with NeuralMesh ran for nine months and became core infrastructure, how they checkpoint K2 training runs every 30 minutes without breaking a sweat, and how they sustain 30 to 40 percent Model FLOPs Utilization across their largest pre training runs. Hear from Will Beddow, Gabriel Menezes, and Sangwu Lee on reliability, proactive support, and the thousands of engineering hours they saved by not building custom storage on object. *CHAPTERS* 0:00 Works as advertised out of the box 00:12 What KREA builds 00:47 Every component has to keep up 00:56 Life before WEKA on Ceph 01:12 The 30 day POC that ran nine months 01:30 Training K2 on NeuralMesh 01:58 Up to 50x throughput and IOPs 02:11 Sustaining 30 to 40 percent MFU 02:33 Support that flags problems first 02:59 Thousands of engineering hours saved *Learn more:* WEKA and KREA: https://weka.io/customers/krea-ai//?utm_source=youtube&utm_medium=social&utm_content=krea-page NeuralMesh product page: https://www.weka.io/product/neuralmesh/?utm_source=youtube&utm_medium=social&utm_content=product-page/ Request a demo: https://www.weka.io/lp/product-tour/?utm_source=youtube&utm_medium=social&utm_content=product-tour Subscribe for more AI infrastructure content: https://www.youtube.com/@WekaIO?sub_confirmation=1 About WEKA: WEKA NeuralMesh is the data platform built for AI, delivering the performance and efficiency needed to run agentic and LLM workloads at production scale. #CustomerStory #WEKA #AIInfrastructure #NeuralMesh #FoundationModels