The American open-source AI lab behind Hermes — now with the Psyche decentralized network
HQ: New York, NY, United States | Founded: 2023 | Employees: ~40 (2026, est.) | Stage: Early Stage VC (~$1.5B valuation, in progress) | Website: https://nousresearch.com
Nous Research was founded in 2023 in New York by Karan Malhotra (CEO) and Jeffrey Quesnelle, building on Quesnelle's earlier open-source work fine-tuning models like WizardLM. The lab quickly became one of the most respected names in the open-source AI community: its Hermes fine-tunes — instruction-tuned versions of base models like Llama and Qwen — became the default 'instruct' layer for thousands of derivative models across Hugging Face, effectively the community standard for high-quality open instruct models. DeepHermes extended the family with deep-thinking capabilities.\n\nIn 2024 Nous released DisTrO (Distributed Training Over-the-Internet), a breakthrough optimization that cut inter-node communication during training by as much as 1,000x — the paper that made decentralized training of frontier-scale models look feasible for the first time and earned Nous a 'Nobel-adjacent' reputation in the open-AI community. The lab followed with Nous Forge (a synthetic-data curation and model-optimization service) and, in 2025-2026, the Hermes agent — an open agentic AI product that moved Nous from model-maker to consumer-facing agent company, with the Psyche Network (a decentralized training network) as its long-horizon moonshot.\n\nFinancially, Nous raised roughly $70M across early rounds, and in July 2026 was reported to be raising at least $75M more at a ~$1.5B valuation, led by Robot Ventures with significant USV participation. The round values not just the models but the thesis: that open, decentralized AI — American-style — can compete with closed frontier labs, and that the infrastructure (DisTrO, Psyche) to make it happen is worth building.
Nous Research is the most interesting open-source AI lab in America: a tiny team with outsize influence. Its Hermes fine-tune family became the default 'instruct' layer for thousands of open models, DisTrO showed how to cut distributed-training bandwidth by 1,000x (a potential moonshot for decentralized training), and the Hermes agent turned the lab into a serious agentic-AI contender. The market validated it: ~$75M round in talks at a ~$1.5B valuation (July 2026, led by Robot Ventures with USV). The bull case is open-source gravity plus genuinely novel infrastructure research (DisTrO, Psyche); the bear case is that open-weights labs monetize poorly, frontier labs keep pulling away, and the decentralized-training thesis is years from proven economics.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki