Slingshot AI trains models for Ash, its mental-health assistant, on Nebius GPU clusters. The company acknowledges working with Nebius to develop a foundation model for psychology. Models used across a conversation range from 32 billion to 235 billion parameters, including a Qwen3-235B backbone. DeepSpeed and ZeRO-3 distribute training across GPUs. SkyPilot automates training runs and reduces developers’ need to manage remote GPU servers.
Fine-tuning and reinforcement learning use expert-guided reward models. The customer story describes blending these with user engagement signals and LLM-based evaluations, giving expert rewards the greatest weight. Nebius’s healthcare page says 50,000 beta users helped train and refine Ash over 18 months. Ash’s safety guardrails use a fast input classifier followed by a safety-tuned LLM for flagged inputs. A clinician reviews every flagged session. Nebius’s healthcare page also cites an NYU study that found Ash helps users feel more connected.
Ash launched in July 2025 as a free Android and iOS app. Nebius’s customer story describes ongoing retraining, reinforcement learning and inference for some models. Managed Kubernetes for Ash inference remains a stated next step to automate scaling and handle fluctuating user loads.
Nebius says Slingshot AI trained Ash’s 32B–235B models on Nebius GPU clusters using DeepSpeed and Zero-3, with expert-guided reward models for fine-tuning and reinforcement learning. The page also cites an NYU study that found Ash helps users feel more connected.
Cloud for Scientific AI and Healthcare8 Oct 2026 · nebius.com · Web pageNebius says Slingshot AI trained Ash’s 32B–235B models on Nebius GPU clusters using DeepSpeed and Zero-3, with expert-guided reward models for fine-tuning and reinforcement learning. The page also cites an NYU study that found Ash helps users feel more connected.
Nebius Customer story
28 Aug 2026 · Nebius · Web page
Nebius describes Slingshot AI training Ash’s models on its GPU clusters and planning managed Kubernetes for inference. The story also describes Ash’s two-pass safety guardrails, using an input classifier and a safety-tuned LLM, with clinician review of every flagged session.
Nebius Customer story28 Aug 2026 · Nebius · Web pageNebius describes Slingshot AI training Ash’s models on its GPU clusters and planning managed Kubernetes for inference. The story also describes Ash’s two-pass safety guardrails, using an input classifier and a safety-tuned LLM, with clinician review of every flagged session.
Slingshot acknowledges Nebius's role in its psychology foundation-model work
10 Dec 2025 · Slingshot AI · Web page
Slingshot AI says it is working with Nebius to build a foundation model for psychology. The shared Nebius post names Slingshot AI as a customer and describes Ash’s training, fine-tuning and inference workloads through their collaboration.
Slingshot acknowledges Nebius's role in its psychology foundation-model work10 Dec 2025 · Slingshot AI · Web pageSlingshot AI says it is working with Nebius to build a foundation model for psychology. The shared Nebius post names Slingshot AI as a customer and describes Ash’s training, fine-tuning and inference workloads through their collaboration.
Slingshot AI develops specialized AI models for mental health and makes Ash, a consumer app offering emotional support through voice and text conversations. Ash is designed for people working through everyday challenges, stress, difficult feelings, and relationship concerns. It remembers previous conversations, identifies patterns across thoughts and behaviors, and provides weekly insights and personalized next steps. The company combines machine learning research with clinical expertise to train its psychology-focused model on behavioral health interactions and multiple therapeutic approaches.
Slingshot releases 12-month follow-up results for Ash
The Hemingway Report covered updated research reporting sustained reductions in depression and anxiety among Ash study participants at 12 months. Only 62% of the initial cohort completed the follow-up, and the study had no control group, so it cannot establish that Ash caused the improvements.
Ash withdrawal survey published in JMIR Medical Informatics
Slingshot AI researchers and a Northwestern collaborator published a survey of 393 former UK Ash users following its January withdrawal. Among respondents answering the primary-support question, 27.1% relied most on general-purpose AI and 26% sought no support. The survey had a 5.3% response rate and first appeared as a preprint in March.
Slingshot AI publishes one-year results from Ash study
Slingshot AI reported sustained improvements in depression and anxiety scores in its observational Ash study conducted with New York University. Of 299 participants at baseline, 185 completed the one-year assessment. The results appear in an updated preprint.
Slingshot AI trains models for Ash, its mental-health assistant, on Nebius GPU clusters. The company acknowledges working with Nebius to develop a foundation model for psychology. Models used across a conversation range from 32 billion to 235 billion parameters, including a Qwen3-235B backbone. DeepSpeed and ZeRO-3 distribute training across GPUs. SkyPilot automates training runs and reduces developers’ need to manage remote GPU servers.
Fine-tuning and reinforcement learning use expert-guided reward models. The customer story describes blending these with user engagement signals and LLM-based evaluations, giving expert rewards the greatest weight. Nebius’s healthcare page says 50,000 beta users helped train and refine Ash over 18 months. Ash’s safety guardrails use a fast input classifier followed by a safety-tuned LLM for flagged inputs. A clinician reviews every flagged session. Nebius’s healthcare page also cites an NYU study that found Ash helps users feel more connected.
Ash launched in July 2025 as a free Android and iOS app. Nebius’s customer story describes ongoing retraining, reinforcement learning and inference for some models. Managed Kubernetes for Ash inference remains a stated next step to automate scaling and handle fluctuating user loads.
Nebius says Slingshot AI trained Ash’s 32B–235B models on Nebius GPU clusters using DeepSpeed and Zero-3, with expert-guided reward models for fine-tuning and reinforcement learning. The page also cites an NYU study that found Ash helps users feel more connected.
Cloud for Scientific AI and Healthcare8 Oct 2026 · nebius.com · Web pageNebius says Slingshot AI trained Ash’s 32B–235B models on Nebius GPU clusters using DeepSpeed and Zero-3, with expert-guided reward models for fine-tuning and reinforcement learning. The page also cites an NYU study that found Ash helps users feel more connected.
Nebius Customer story
28 Aug 2026 · Nebius · Web page
Nebius describes Slingshot AI training Ash’s models on its GPU clusters and planning managed Kubernetes for inference. The story also describes Ash’s two-pass safety guardrails, using an input classifier and a safety-tuned LLM, with clinician review of every flagged session.
Nebius Customer story28 Aug 2026 · Nebius · Web pageNebius describes Slingshot AI training Ash’s models on its GPU clusters and planning managed Kubernetes for inference. The story also describes Ash’s two-pass safety guardrails, using an input classifier and a safety-tuned LLM, with clinician review of every flagged session.
Slingshot acknowledges Nebius's role in its psychology foundation-model work
10 Dec 2025 · Slingshot AI · Web page
Slingshot AI says it is working with Nebius to build a foundation model for psychology. The shared Nebius post names Slingshot AI as a customer and describes Ash’s training, fine-tuning and inference workloads through their collaboration.
Slingshot acknowledges Nebius's role in its psychology foundation-model work10 Dec 2025 · Slingshot AI · Web pageSlingshot AI says it is working with Nebius to build a foundation model for psychology. The shared Nebius post names Slingshot AI as a customer and describes Ash’s training, fine-tuning and inference workloads through their collaboration.
Slingshot AI develops specialized AI models for mental health and makes Ash, a consumer app offering emotional support through voice and text conversations. Ash is designed for people working through everyday challenges, stress, difficult feelings, and relationship concerns. It remembers previous conversations, identifies patterns across thoughts and behaviors, and provides weekly insights and personalized next steps. The company combines machine learning research with clinical expertise to train its psychology-focused model on behavioral health interactions and multiple therapeutic approaches.
Slingshot releases 12-month follow-up results for Ash
The Hemingway Report covered updated research reporting sustained reductions in depression and anxiety among Ash study participants at 12 months. Only 62% of the initial cohort completed the follow-up, and the study had no control group, so it cannot establish that Ash caused the improvements.
Ash withdrawal survey published in JMIR Medical Informatics
Slingshot AI researchers and a Northwestern collaborator published a survey of 393 former UK Ash users following its January withdrawal. Among respondents answering the primary-support question, 27.1% relied most on general-purpose AI and 26% sought no support. The survey had a 5.3% response rate and first appeared as a preprint in March.
Slingshot AI publishes one-year results from Ash study
Slingshot AI reported sustained improvements in depression and anxiety scores in its observational Ash study conducted with New York University. Of 299 participants at baseline, 185 completed the one-year assessment. The results appear in an updated preprint.
