How they use Nebius

Outerbounds runs Metaflow workflows and AI experiments on Nebius GPU compute. In its joint webinar, data scientist Eddie Mattia describes his team’s experience, followed by LLM fine-tuning and post-training examples. Outerbounds also offers a bring-your-own-cloud integration that runs in customers’ cloud accounts. The integration supports executing a Metaflow workflow on Nebius by adding @nebius, optionally running parts in existing cloud environments. Outerbounds automatically containerizes the project in seconds. Metaflow’s @checkpoint integration connects to Nebius Object Storage. Nebius infrastructure monitoring provides workflow metrics, alongside cross-cloud cost optimization tools in the Outerbounds UI. In March 2025, the partners outlined plans for additional GPU support, more managed services and further integration improvements.

Outerbounds used on-demand Nebius H100 compute pools for autonomous-inference benchmarks and tested Nebius AI Studio endpoints, including Qwen QwQ 32B. Its workload-aware Qwen2.5 72B benchmark completed over seven times faster than the compared Together.AI endpoint. In its Qwen QwQ 32B benchmark, Nebius AI Studio and Together.AI performed similarly, with different latency spreads.

NVIDIA also documented Outerbounds’ Reddit Recommender demo actively deployed on a Nebius H100 compute pool. The deployment used NVIDIA NIM MessageFormatter and Embeddings models. NVIDIA also describes GPU-accelerated vector indexing and reports that the single-H100 implementation on Nebius was over twice as fast as the compared CPU instance at half the cost.