CubeStudio
An open-source, cloud-native, one-stop AI platform covering traditional machine learning, deep learning and large-model workflows, from multi-tenant compute scheduling through pipeline orchestration to model deployment and an LLMOps agent layer.
CubeStudio is an open-source, cloud-native AI platform covering the full pipeline from traditional machine learning and deep learning through large-model training and serving (MLOps, MaaS, compute scheduling and training/inference platforms). It is MIT licensed and free for commercial use, and the open-source version reports thousands of private, self-hosted deployments at companies.
It provides multi-tenant management with compute allocation and a compute rental marketplace with token relay, drag-and-drop pipeline orchestration, distributed multi-machine multi-GPU training, hyperparameter search, inference serving, vGPU virtualization, cloud-edge-device coordination and edge computing, and automated labeling across image, text and audio data. Large-model features include supervised fine-tuning, reward-model and reinforcement-learning training, multi-machine inference through vLLM, Ollama and MindIE, a private knowledge base and LLMOps agent layer, and a model marketplace.
It natively supports domestic Chinese accelerator hardware such as Ascend, Cambricon, Hygon and Moore Threads alongside x86 and ARM CPUs, plus IB/RoCE/RDMA high-speed networking, for offline, private deployment.