Company Profile / Open-Source AI

Nous Research

The open-source AI lab building frontier-grade models - and training them across a swarm of GPUs scattered around the internet.

Founded 2023 Austin & New York, US Backed by Paradigm

Nous Research, the open-source lab whose Hermes models have been downloaded tens of millions of times, photographed through its own emblem - the face of a company that started as a Discord server and now trains AI without a data center.

The Dispatch

An open-source lab betting against the data center

Nous Research is an artificial-intelligence lab with an unusual thesis: the most powerful AI models should not be locked inside a handful of corporations, and they should not require a single billion-dollar data center to train. Founded in 2023 out of a community of open-source researchers, the company builds open-weight language models - its Hermes family - and the infrastructure to train them across ordinary hardware spread around the world.

Where labs like OpenAI and Anthropic keep both their weights and their training methods private, Nous publishes model weights openly on Hugging Face and pairs them with detailed technical reports. The pitch to developers is straightforward: download the model, inspect it, fine-tune it, and run it however you like. That openness turned Hermes into a default building block for the open-source AI ecosystem, with a reported 33 million-plus downloads.

The second half of the thesis is decentralization. Through its Psyche network, Nous coordinates heterogeneous compute - from a gamer's RTX 4090 to a data center's H100 - into a single, fault-tolerant training run coordinated on the Solana blockchain. It is an attempt to break the assumption that frontier AI is only for those who can afford the largest clusters.

Products & Services

What Nous builds

Models · since 2023

Hermes family

Open-weight, fine-tuned language models known for steerability, tool use, and reasoning. The line runs from Hermes 2 through Hermes 4.

Release · 2025

Hermes 4

Open-weight 14B, 70B and 405B models with hybrid reasoning - toggle explicit <think> chains on for hard problems, off for quick replies.

Network · 2025

Psyche

A decentralized training network on Solana that stitches consumer and datacenter GPUs into one fault-tolerant training run.

Optimizer · 2024

DisTrO

Distributed training optimizers that slash inter-GPU communication, making it practical to train over ordinary internet connections.

Interface · 2025

Nous Chat

A hosted chat front end offering access to Nous models, including free trial access to Hermes 4.

Platform · 2026

Hermes Agent & Portal

An agentic product and access platform reported to anchor the company's 2026 funding round.

Who uses it

The customers

Nous serves open-source AI developers, researchers, and startups that want to build on models they can actually own and modify. Hermes has become a common base for fine-tuning in the Hugging Face community, downloaded tens of millions of times. On the compute side, Psyche draws in Solana and crypto-ecosystem participants who contribute idle GPUs to distributed training runs.

The problem

What it solves

Two bottlenecks define modern AI: access and cost. Closed models leave developers dependent on APIs they cannot inspect or steer. Training frontier models demands data centers few can afford. Nous attacks both - open weights remove the black box, and Psyche's decentralized training aims to remove the data-center requirement, with the training process made transparent enough to verify.

How it's different

Against the grain

Open, not closed

Unlike OpenAI or Anthropic, Nous ships open weights and technical reports so anyone can inspect, run, and modify its models.

Distributed, not centralized

Where peers depend on single mega-clusters, Psyche trains across scattered GPUs coordinated on-chain.

Community, not corporate

Nous grew from a research community, and participation - in models and compute - remains central to how it works.

Business model

How it makes money

Nous gives models away to drive adoption, then monetizes around them: hosted inference through Nous Chat and Nous Portal, enterprise engagements, and a token-incentivized compute network in Psyche. Venture and crypto-native capital funds the research engine while the open models build distribution.

Expertise

Where it's deep

The lab's core strengths are post-training and fine-tuning, model behavior and alignment, reasoning, and the systems research behind distributed training - including the DisTrO optimizers that make training over the internet feasible. Hermes 4's benchmark results reflect that focus on squeezing frontier performance from open checkpoints.

Funding & Market

The money

Seed (2024)~$20M total

Distributed Global · OSS Capital · Delphi

Series A (Apr 2025)$50M · $1B valuation

Led by Paradigm · Together AI

New round (2026, reported)$75M+ · ~$1.5B valuation

Robot Ventures · Union Square Ventures

In a market split between closed frontier labs (OpenAI, Anthropic, Google DeepMind) and open-weight players (Meta Llama, Mistral, DeepSeek), Nous occupies a distinct corner: open weights plus decentralized training. Its nearest peers in distributed training include Prime Intellect and Gensyn.

The Record

Timeline

2023

Nous Research is founded

A community of open-source researchers - Jeffrey Quesnelle, Karan Malhotra, Teknium and Shivani Mitra - formalizes into a lab and begins releasing Hermes models.

2024

Seed funding and DisTrO

Closes a $5.2M seed (later ~$20M total) and publishes distributed-training optimizers that enable training over ordinary internet connections.

2025

Series A and Psyche launch

Raises $50M led by Paradigm at a $1B valuation and launches the Psyche decentralized training network on Solana.

2025

Hermes 4 released

Ships open-weight 14B/70B/405B models with hybrid reasoning and a 94-page technical report, with strong math benchmark results.

2026

$1.5B valuation talks

Reported to be finalizing a $75M+ round at a ~$1.5B valuation centered on its Hermes Agent.

The Founders

Who's behind it

Co-founder & CEO

Jeffrey Quesnelle

Leads the company and its research direction toward open, unrestricted models.

Co-founder · Head of Behavior

Karan Malhotra

Focuses on how the models behave, align, and respond.

Co-founder · Head of Post-Training

Teknium

Drives the fine-tuning and post-training work that defines the Hermes line.

Co-founder

Shivani Mitra

Part of the founding team that turned a research community into a company.

Questions

FAQ

What does Nous Research do?+

It is an open-source AI lab that builds open-weight language models (the Hermes family) and decentralized infrastructure (Psyche) for training them across distributed compute.

Who founded Nous Research and when?+

It was founded in 2023 by Jeffrey Quesnelle (CEO), Karan Malhotra, the researcher known as Teknium, and Shivani Mitra, growing out of an open-source AI community.

What is the Psyche network?+

Psyche is a decentralized training network built on the Solana blockchain that coordinates heterogeneous GPUs - from consumer RTX 4090s to datacenter H100s - into fault-tolerant training runs.

How much has Nous raised?+

Roughly $20M in seed rounds (2024) and a $50M Series A led by Paradigm at a $1B valuation (April 2025); in July 2026 it was reported to be finalizing a $75M+ round at a ~$1.5B valuation.

What makes Hermes 4 different?+

Hermes 4 is a family of open-weight models with hybrid reasoning that can toggle explicit reasoning on or off, released with open weights and without baked-in content restrictions.