How they use Nebius AI Cloud

Recraft trains image models and runs inference on Nebius. It began using the platform during technical preview and trained its earlier 20-billion-parameter image model from scratch. Its V4 model used Nebius B200 GPUs for training and serving. Nebius reports that Recraft deployed B200 training and inference clusters in less than a day with minimal code changes and engineering support. Recraft has also described moving parts of its pipeline from Hopper to Blackwell hardware and testing B300 GPUs. The V4 story records earlier Ampere deployments.

A July 2026 interview describes inference across several data centres, with different GPUs handling different parts of the image-generation pipeline. In an earlier training run for a 20-billion-parameter model, Nebius says it fixed a network bottleneck that made training six times faster. A technical recording describes the team's use of managed Kubernetes, Kubeflow and InfiniBand for model development.

The earlier customer story also lists Nebius Object Storage. Managed Kubernetes handled nodes and GPU drivers, while Kubeflow allocated training resources. Recraft added training-step heartbeat alerts as Nebius addressed hardware and network failures. For V4, Nebius supplied Blackwell GPUs connected by Quantum-2 InfiniBand. Recraft changed part of its PyTorch stack from TensorRT compilation to torch.compile. Nebius reports automated replacement of failed nodes and training restarts from the last checkpoint. Recraft primarily uses bfloat16 for training and FP8 for its frozen text encoder and LLM inference. It is experimenting with NVFP4 to improve inference speed. Recraft’s subprocessor list names Nebius Group N.V. for cloud infrastructure services in the United States and EU.