Vibeleaderboard
Nous Research

model publisher

Also indexed as nous · nous-research

Nous Research

Nous Research matters because it contributed an adopted context-extension method, helped make community post-training a durable layer of the open-model ecosystem, and then pushed beyond fine-tuning into low-bandwidth training and local agents. Its work tests whether an organization can combine open weights, distributed infrastructure, and a commercial product without collapsing those distinct claims into one open-source label.2,3,5,1

Official site ↗Models in the Atlas →

Profile

Overview

A company with community roots

Nous Research began as an informal online research community in 2022 and was formalized as an AI company in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium, and Shivani Mitra. It emerged from a community of model trainers and researchers rather than from a single university or larger technology company. The organization is best known for open-weight post-training, but it now spans language models, a local agent product, and infrastructure for distributing model training across less tightly connected machines.1,5,8

Hermes as post-training research

Nous researchers first established a broader research record with YaRN, a method for extending a pretrained language model's context window more efficiently than full-length fine-tuning. Hermes later became Nous's main model line. The releases post-train openly available base models for instruction following, roleplay, tool use, and reasoning rather than pretraining every generation from scratch. The Hermes 4 report documents data curation, synthetic-data generation, hybrid reasoning, evaluation across math, code, knowledge, comprehension, and alignment tasks, and public release of the weights. Technical credit still needs to distinguish Nous's post-training from the underlying base-model architecture.3,2

Training across the internet

Nous also developed DisTrO, short for Distributed Training Over-The-Internet, as a family of optimizers intended to reduce the communication burden between accelerators. A preliminary report and public training demonstration explored whether large-model training could work across slower and more geographically dispersed links. Psyche extends that research into a coordination network for decentralized runs. Independent analysis recognizes the technical ambition while noting that public observability and governance remain less mature than the open-source code and released weights.4,5,6

From open models to an agent business

In 2026 the company expanded from models and training research into Hermes Agent, a local agent with built-in skills, messaging integrations, and a hosted option. TechCrunch reported in July that Nous was seeking at least $75 million at a $1.5 billion valuation and identified the four founders. That financing report is evidence of investor demand, not proof of model leadership. It raises the execution stakes for converting community credibility and open releases into a sustainable product without obscuring which components are open, hosted, or venture-controlled.1,7

Notable contributions

  1. 01YaRN context-window extensionYaRN introduced a compute-efficient method for extending a pretrained language model's usable context window through changes to rotary position embeddings and a staged training recipe.3
  2. 02Low-bandwidth distributed optimizationDisTrO explored aggressive reduction of inter-accelerator communication and demonstrated a public 15-billion-parameter training run as a step toward training across ordinary network links.4,5
  3. 03Psyche decentralized training infrastructurePsyche turned Nous's optimizer research into a network and coordination system for geographically distributed model-training runs. The project has released code and model artifacts, while independent review still identifies gaps in public network observability and governance.6,5
  4. 04Hermes Agent's learned skills workflowHermes Agent packages local execution, messaging, built-in skills, and the ability to turn successful task patterns into reusable skills as a distinct product from the Hermes language models.1,7