cube-studio

Why we included this project

CubeStudio is a broad, Kubernetes-based platform that tries to cover the whole ML lifecycle, and for teams that want one self-hosted system rather than a patchwork of tools, that ambition pays off. It handles the usual MLOps duties like multi-tenant user and project management, a drag-and-drop pipeline editor, distributed training, hyperparameter search, and serving, but it also goes deep into large-model work: SFT fine-tuning, reward-model and RL training, multi-node inference through vLLM, Ollama, or MindIE, and a private knowledge-base and agent layer. The platform was built with Chinese enterprise needs in mind, including native support for domestic accelerators like Ascend, Cambricon, Hygon, Moore Threads, and Muxi, plus RDMA networking and air-gapped, offline deployment. That makes it a realistic candidate for organizations running on national stack hardware, whether they are doing classical ML or LLM fine-tuning in-house.