Nous Research

Nous Research develops open-source language models and infrastructure for distributed training. Its work includes the Hermes model family, the Hermes agent, and DisTrO, a project aimed at reducing communication between GPUs during training. Nous describes model architecture, data synthesis, fine-tuning and reasoning as its primary research areas, and unrestricted model availability and use as part of its mission.[2][5][6][4][11]

By July 2026, Nous had expanded from language models into an open-source agent product. Hermes shipped with skills for web search, coding and image understanding, and was designed to learn from usage and build additional skills automatically. It could automate tasks, communicate through Telegram and Discord, and run on a desktop or virtual private server. Nous also offered a cloud-hosted version; TechCrunch reported monthly tiers of $20–$200 and roughly 214,000 GitHub stars with nearly 40,000 forks. Those repository figures measure attention to the project, rather than users or paying customers.[11]

The agent followed earlier work on Hermes instruction models and experiments with language-model interfaces. That history begins with a volunteer research group and includes WorldSim, a prompt-driven simulation demonstrated in 2024.[10][5][1][11]

Cofounder Karan Malhotra described Nous as an open-source research organization created in 2022, when volunteers met through Discord, GitHub and Twitter. They adapted existing models such as Meta’s Llama and Mistral and released versions under the Hermes name. TechCrunch separately dates the company’s founding to 2023 and names Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium and Shivani Mitra as founders.[10][11]

WorldSim, introduced at the end of March 2024, used a Claude prompt to create an imagined world navigated through invented shell commands. In a demonstration recorded by Swyx at Replicate HQ and published in Latent Space’s April 27 episode, Malhotra explained how he mimicked a command-line interface and wrapped interactions in XML tags. Commands such as cd and ls elicited simulated directories and files, rather than operating a real filesystem.[1]

In the demonstration, Malhotra created a universe and then Twitter. Because he had not specified the timing, the simulation placed Twitter at the universe’s creation and introduced human users much later. The example showed how prompt constraints shaped the generated story; Malhotra cautioned, “So, I don't expect these numbers to be right.” WorldSim also inspired the separate WebSim hackathon project. Rob Haisfield recalled that Karan introduced him to his cofounder Shawn.[1]

The Hermes 3 technical report, submitted on August 15, 2024 by Ryan Teknium, Jeffrey Quesnelle and Chen Guang, presented a generalist instruction-following and tool-use model with reasoning and creative abilities. The authors described it as “neutrally-aligned” and reported that its largest version, Hermes 3 405B, achieved state-of-the-art performance among open-weight models on several public benchmarks at publication.[5]

Hermes 4 combined structured, multi-turn reasoning with broad instruction following, with model weights made public. Its report was submitted on August 25, 2025 and revised on September 2. The author team comprised Ryan Teknium, Roger Jin, Jai Suphavadeeprasit, Dakota Mahan, Jeffrey Quesnelle, Joe Li, Chen Guang, Shannon Sands and Karan Malhotra.[6]

Earlier work by overlapping researchers addressed context length. [YaRN](https://arxiv.org/abs/2309.00071), initially submitted on August 31, 2023 by Bowen Peng, Jeffrey Quesnelle, Honglu Fan and Enrico Shippole, extends context windows in models using rotary position embeddings, or RoPE. Its February 2026 revision reports using ten times fewer tokens and 2.5 times fewer training steps than previous context-extension methods, and extrapolating beyond the context lengths in a fine-tuning dataset.[8]

Nous’s [DisTrO](https://github.com/NousResearch/DisTrO)—Distributed Training Over-The-Internet—targets the communication demands of training across GPUs. Its project chronology dates a preliminary report to August 26, 2024, a 15B-model training announcement to December 2, and Psyche Network and Nous Consilience 40B entries to May 14, 2025.[4]

The DisTrO project identifies [DeMo](https://arxiv.org/abs/2411.19870) as its seed research. DeMo decouples local momentum updates, applies a fast orthonormal transform followed by top-k sparsification, and reuses the momentum buffer for error feedback. Its authors are Bowen Peng, Lizhang Chen, Baiyu Su, Jeffrey Quesnelle, Diederik P. Kingma and Qiang Liu. The February 2026 revision reports up to 85 times less transmitted data per GPU than AdamW-DDP in experiments with 300M- and 1B-parameter language models, while achieving comparable loss and accuracy—a communication result, rather than a training-time speedup.[9][4]

[Token-Superposition Training](https://arxiv.org/abs/2605.06546), presented in May 2026 by Bowen Peng, Théo Gigant and Jeffrey Quesnelle, addresses pretraining throughput. It first combines contiguous tokens into bags and trains with multi-hot cross-entropy, then returns to standard training for recovery. The authors evaluated 270M- and 600M-parameter models and validated the method at 3B and on a 10B-A1B mixture-of-experts model. For the latter configuration, they reported up to a 2.5-fold reduction in total pretraining time at equal loss.[7]

In April 2025, Nous raised a $50 million Series A, financed almost entirely by Paradigm, which confirmed the amount to Fortune. Fortune reported approximately $20 million in earlier seed rounds from investors including Distributed Global, North Island Ventures and Delphi Ventures. An unnamed source supplied a $1 billion token valuation for the Series A. At the time, Nous had a roughly 20-person team, with much of the new funding directed toward compute and expanded research capabilities.[10]

Malhotra explained the decision to use crypto incentives for distributed compute as “less as a donation but more as a transaction.” In April 2025, Nous planned to launch its distributed training system on Solana, initially allowing participation by people with access to data centers. The reward design remained undecided between a proprietary token and Solana’s native cryptocurrency. Malhotra also argued that blockchain could discourage malicious contributions, while Fortune noted unresolved questions about preventing data poisoning without sacrificing training efficiency.[10]

In July 2026, TechCrunch reported that Nous was finalizing at least $75 million in new financing led by Robot Ventures, with significant USV participation, at a $1.5 billion valuation. The terms came from three unnamed sources; Nous declined to comment, and the named investors did not respond. The report described a round in progress, rather than a closed financing.[11]