Nous Research

Nous Research is an American artificial-intelligence laboratory that develops open-weight language models and an open-source personal agent, Hermes Agent, on the premise that widely available AI should be inspectable, auditable and controlled by its users rather than by a platform.^[1][3]

History

Nous was formed in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium and Shivani Mitra — a small team without institutional backing that worked in the open. Its early contribution was post-training: fine-tuning Meta's Llama models with an emphasis on user-alignment, meaning the model follows the worldview and instructions of the individual user rather than a policy set centrally by a company. The group published work on synthetic data and helped define open-source tool-calling conventions, establishing that a small team could produce fine-tuned models competitive with far better-funded laboratories.^[1]

Throughout 2025 the team built an agent stack around its models and released Hermes Agent under an MIT licence in February 2026: a tool that chooses where it runs and which models it uses, remembers projects, and writes reusable skills as it works so knowledge carries across sessions. In October 2026 Nous confirmed a $90 million Series B at a $1.5 billion valuation led by Robot Ventures, with NVIDIA, Union Square Ventures, Menlo Ventures, Samsung, Microsoft's M12, Y Combinator and others participating, bringing total funding to about $158 million; the capital was earmarked for Hermes for Businesses, a push to bring customised, multi-step agents to corporate users.^[1][2][4]

Key Concepts and Technology

Three commitments define the work. Open weights — models are published so anyone can download, run, fine-tune and audit them, rather than accessed only through an API. Post-training focus — Nous competes on instruction-following, tool use, coding and reasoning after pre-training rather than on raw scale. Local-first agents — the agent runs where the user decides, on hardware the user controls, with models the user selects.^[1]

This matters for safety arguments as much as for deployment. Nous's stated position is that openness makes systems safer through auditability rather than through gatekeeping: a model whose weights and inference can be examined can be tested, patched and constrained by anyone, whereas a closed system must be trusted on the provider's word.^[1] The commercial counterweight is that open-weight models still cost money to serve at scale, which is why the same release train typically carries both an openly downloadable model and a hosted API.

Applications and Impact

Open-weight models of this class are used for private document analysis, on-premise assistants, code assistants in regulated environments, research fine-tuning, and agents that must run without sending data to a third-party cloud. Nous helped normalise the pattern of small laboratories shipping competitive open releases, a trend that by 2026 sees open-weight drops from several major labs each month. The company also illustrates the sector's economics: a volunteer-adjacent open-source project converted into a billion-dollar venture-backed business by making the agent — not the model weights — the product.

>See Also

• Open-Weight Models
• Llama
• AI Alignment
• Constitutional AI
• Small Language Models
• Malaysia AI Governance Framework
• Sovereign AI