Case Study: Sword Health achieves faster, safer AI mental health support with Nebius

Sword Health scales Dawn to 200B+ on Nebius, cutting tail latency from 20s to under 12s

Sword Health, an AI care company, faced the challenge of scaling its new direct-to-consumer mental health wellbeing AI, Dawn, to handle long, safety-sensitive conversations. Their goal was to move from a 30B parameter model to a much larger one for higher quality, but this risked making production inference unbearably slow and costly. They partnered with vendor Nebius to find a solution that could meet their strict clinical safety and latency requirements.

Nebius implemented a solution using custom speculative decoding on dedicated NVIDIA Blackwell endpoints. This allowed Sword Health's Dawn AI to successfully jump from a 30B model to a 200B+ parameter model. The solution dramatically cut tail latency from over 20 seconds to under 12 seconds in production, enabling high-quality, responsive mental health support at scale without compromising on safety or performance.

Sword Health