How they use Nebius

Sword Health uses Nebius for training and inference across Dawn, its mental-health product, and Thrive, its AI-guided physical-care platform. Nebius says Dawn expanded Sword Health's AI Care model into direct-to-consumer mental wellbeing support, combining years of clinical AI development with new inference architecture. Nebius says Dawn uses engagement over time for higher-quality care. Ricardo Rei describes Dawn as available to anyone, whenever they need it. A Sword Health engineer says the team moved training infrastructure to Nebius and trained speculative decoders deployed through Token Factory. Sword Health's own interview describes building Dawn's foundation model with Nebius as its cloud provider. Nebius's customer story says Sword Health migrated Thrive's movement-analysis vision models to Nebius cloud. Thrive delivered approximately 4 million AI-guided sessions in 2025. Nebius's healthcare page separately reports that Sword Health delivered 10 million AI-guided sessions in 2025.

Nebius reports scaling Dawn from a 30B model to more than 200B parameters. The initial 30B model did not meet Sword Health's quality bar. Nebius says Dawn screens every user message for distress. Sequential MindGuard screening added latency, so the team treated the non-streaming response and cascaded models as one latency budget. Its customer story describes production deployment within a month through Token Factory on dedicated NVIDIA Blackwell endpoints. Custom speculative decoding achieved a draft-model acceptance rate above 50%. KV caching and prefix-aware routing supported guardrail screening and multi-step reasoning, with the setup running in BF16. Dawn targeted long, multi-session conversations with safety, governance and cost requirements. Nebius says the resulting platform cleared Sword Health's quality bar.

Nebius's May 2026 post reports production tail latency falling from over 20 seconds to under 12 seconds. At Inflection 2026, Nebius reported the same reduction for end-to-end response latency in the setup it presented. The customer story describes a P99 latency budget below 12 seconds for non-streaming responses. Nebius's customer listing describes production conversations supported by Nebius AI Cloud, MindGuard safety guardrails and the MindEval evaluation framework.