How they use Nebius

Higgsfield uses Nebius for image and video model training and inference. CEO Alex Mashrabov says the company trained Soul 2.0 on Nebius B300 GPUs. Earlier work covered diffusion models for image editing and keyframe generation. Nebius supported streaming dataset storage and distribution. Image-caption pairs were filtered, encoded into latents and cached by resolution in a shared, location-aware store. Training nodes pulled preprocessed batches asynchronously. Continuous preference-driven training loops used Direct Preference Optimization with human-ranked outputs. NVIDIA's account also describes NVSentinel and DCGM for infrastructure reliability, plus CUDA libraries and InfiniBand networking for creative control and performance.

The company also uses Nebius Token Factory for inference. At Nebius Inflection in June 2026, Nebius described managed serving and autoscaling for Higgsfield's media generation. It reported 4.5 million generations a day at the time. In an earlier interview, Mashrabov discussed choosing GPU configurations for different jobs and testing AI Studio for a video-script workflow. He cited transparent pricing and startup-friendly terms as reasons for choosing Nebius. He also outlined possible future collaboration on serverless media inference and video-processing frameworks.

For Hell Grind, its completed 95-minute AI film, Higgsfield scaled overnight generation to hundreds of Blackwell GPUs on Nebius AI Cloud, then scaled them back down by morning. Joshua Liss described 100,000 unique video generations during the two-week production. In July 2026, Nebius reported around 30 billion tokens per day through AI Cloud and over 20% better inference performance from a custom configuration. A separate account describes co-engineered distillation that reduced inference time and serving cost roughly twentyfold. Nebius's vision page reports 40% cost efficiency gains for production video inference versus alternative providers. Nebius's Q2 2026 shareholder letter reports that Higgsfield expanded its platform usage more than twentyfold since its first contract.