Inherent
Inherent develops AI models and scientific agents for research and discovery. · Model developers
How they use Nebius AI Cloud
Inherent trained Faraday, its 27B-parameter AI Scientist, on Nebius AI Cloud, according to Nebius EMEA General Manager Gary Tierney. Inherent describes Faraday as a model trained through long-horizon reinforcement learning to replicate scientific research using coding agents as tools.
For training, Inherent introduced Replica, a task space initially comprising 310 tasks from 100 machine learning and AI-for-science papers. Tasks ask agents to reproduce research figures within time and compute limits without seeing the original plot. Inherent reports that Faraday produced more faithful replications than its Claude Opus 4.8 and GPT-5.5 baselines across the research categories tested.
Sources
3 sourcesNebius NVIDIA GTC Berlin 2026
19 Oct 2026 · Nebius · Web page
Nebius lists Inherent Labs as a cohost with Nebius and NVIDIA of a breakfast scheduled for October 21, 2026. The page also identifies Gary Tierney as Nebius's EMEA General Manager.
Open original sourceBuilding Agent-First Infrastructure
19 Oct 2026 · Luma · Web page
The Nebius Events listing names Nebius, Inherent and NVIDIA as hosts of Building Agent-First Infrastructure.
Open original sourceCompany context
inherentlabs.aiInherent is a London-based AI research lab developing models and scientific agents for researchers and R&D teams. Its Faraday agent uses reinforcement learning and coding tools to reproduce findings from scientific papers, with the longer-term aim of supporting discovery across disciplines. The lab also developed Replica, a training task suite based on reproducing research figures in fields including machine learning, materials science and weather forecasting. Inherent combines this technical work with research into how human scientists and AI agents collaborate.
InherentRecent updates
1 updateInherent introduces Faraday AI scientist and Replica training tasks
Inherent introduced Faraday, a 27-billion-parameter AI agent trained to replicate scientific research using coding agents. The company says it outperformed Claude Opus 4.8 and GPT-5.5 on its replication evaluation, which uses 310 tasks drawn from 100 research papers.
Inherent