Nous Research

Nous Research is a New York City-based applied AI research organization and company that builds widely used open-weight language models and decentralized training infrastructure. Its flagship Hermes series of fine-tuned large language models has been downloaded tens of millions of times from Hugging Face, and in April 2025 the company raised a $50 million Series A led by crypto venture firm Paradigm at a $1 billion token valuation. ^[6][8] Since 2025 the company has also operated Nous Portal, a first-party inference and subscription platform, and in 2026 its open-source Hermes Agent became one of the most popular repositories on GitHub. ^[20][22][23] In October 2026 Nous Research announced a $90 million funding round, which TechCrunch reported as a Series B led by Robot Ventures at a $1.5 billion valuation, to build Hermes for Businesses, a version of the agent for companies. ^[31][32] Nous Research focuses on model architecture, data synthesis, post-training, and reasoning, with a growing emphasis on decentralized AI training that lets a single model be trained across hardware distributed over the internet rather than inside one data center.

History and Founding

Nous Research traces its origins to 2022, when a group of AI researchers and engineers began collaborating informally through social media platforms, including Discord, GitHub, and Twitter (now X). The founding members met online and started experimenting with existing open-source AI language models, primarily Meta's LLaMA series. They began releasing fine-tuned versions of these models under the name "Hermes," which quickly gained traction in the open-source community.

The organization was formally incorporated in 2023 by four co-founders:

• Jeffrey Quesnelle (CEO at the time of the 2025 Series A, CTO by 2026) holds an M.S. in Computer Science from the University of Michigan-Dearborn, with an undergraduate degree in Computer Science and Mathematics from Oakland University. ^[6][33] Before founding Nous Research, Quesnelle co-authored the book Automotive Ethernet: The Definitive Guide, published by Intrepid Control Systems in 2015. ^[33]
• Karan Malhotra (Head of Behavior) previously served as a Machine Learning Researcher at Stanford Brain Stimulation Lab. He studied Religion and Philosophy at Emory University and received an Associate of Arts degree from Oxford College of Emory University.
• Teknium (Head of Post-Training at the time of the 2025 Series A; by October 2026 his X profile described him as co-founder and lead engineer of Hermes Agent) is an AI researcher and engineer, credited as Ryan Teknium on the Hermes 4 technical report, who previously worked at Stability AI. ^[6][16][61] Teknium led much of the early fine-tuning work and created the GPTeacher dataset that formed part of the initial Hermes training data.
• Shivani Mitra is a co-founder of the organization.

Nous Research started as an all-volunteer project. Early contributors ranged from former physics Ph.D. holders and mathematicians to biologists and software engineers, all collaborating without pay to advance open-source AI. External investment later enabled the organization to compensate its most dedicated members. Reflecting on the project's ambition, co-founder Karan Malhotra said, "We very much came from a mentality that we want to create and serve the world's best AI." ^[8]

Nous Research, Inc. is a Delaware corporation, ^[60] and its careers page lists a New York office alongside remote positions. ^[59] By October 2026 the company was led by chief executive Dillon Rolnick, who signed its October 2026 fundraising announcement as CEO and whose X profile reads "ceo @nousresearch", while Quesnelle's homepage described him as the company's co-founder and CTO. ^[32][34][33] In that announcement Rolnick wrote that Nous was formed in 2023 "to combat the seemingly inexorable trend dragging AI toward centralization and walled gardens," and that, as "a group of upstarts with few resources," they chose post-training as the layer where they could have the most impact. ^[32]

How is Nous Research funded?

Before its 2026 Series B, Nous Research had raised approximately $65 million across multiple rounds, according to FinSMEs; TechCrunch, citing Crunchbase, put the pre-Series B total at $70 million. ^[7][23] TechCrunch reported that the Series B brought total funding to $158 million. ^[31]

The $5.2 million seed round closed on January 10, 2024, co-led by Distributed Global and OSS Capital. ^[15] The Series A round, announced on April 25, 2025 and led by crypto venture capital firm Paradigm, valued Nous Research at a token valuation of $1 billion, with the round financed almost entirely by Paradigm. ^[6][8] The investment reflected growing interest in the intersection of blockchain technology and AI, particularly around decentralized training infrastructure. In July 2026, TechCrunch reported that Nous Research was raising at least $75 million in new funding at a $1.5 billion valuation, in a round led by Robot Ventures with participation from Union Square Ventures and other investors, on the strength of its Hermes Agent product; TechCrunch put the company's total prior funding at $70 million. ^[23]

On October 7, 2026, Nous Research announced that it had raised $90 million. A note on the company's website signed by CEO Dillon Rolnick named NVIDIA, Microsoft's M12, Samsung, Robot Ventures, Union Square Ventures, Y Combinator, and Menlo Ventures "among others" as investors, called the sum "modest in modern fundraising," and said the money would go toward building "Hermes for Businesses." The note did not state a valuation. ^[32] TechCrunch reported the same day that the round was a Series B at a $1.5 billion valuation, confirming its July reporting, and that it was led by Robot Ventures with participation from backers including Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital, where Donald Trump Jr. is a partner. ^[31] The Next Web, which credited The Wall Street Journal with first reporting the deal, said the Journal put the valuation at $1.5 billion. ^[35] Citing the Journal, TechCrunch also reported that Nous had annualized revenue of roughly $36 million by mid-September 2026 and expected to pass $100 million before the end of 2026; the company's own note did not give revenue figures. ^[31][32]

Hermes Model Series

The Hermes series is the flagship product line of Nous Research. Each generation has been fine-tuned on top of a different base model, typically from Meta's LLaMA family or other prominent open-source architectures. The models are known for their strong instruction-following capabilities, low hallucination rates, and neutrally-aligned behavior that avoids excessive content filtering.

Hermes 1 (2023)

The original Nous-Hermes-13b was released in mid-2023 as a fine-tune of Meta's LLaMA 13B. It was trained on over 300,000 instructions, consisting primarily of synthetic GPT-4 outputs. Training data sources included GPTeacher, CodeAlpaca, Evol-Instruct Uncensored, GPT4-LLM, Unnatural Instructions, Camel-AI science datasets (Biology, Physics, Chemistry, Math), and Airoboros' GPT-4 dataset. The model was fine-tuned on 8x A100 80GB GPUs over more than 50 hours.

At the time of release, Nous-Hermes-13b ranked first on several popular benchmarks including ARC-Challenge, ARC-Easy, HellaSwag, and OpenBookQA when compared against the GPT4All benchmark suite. The model used a simple Alpaca-style instruction/response prompt format.

Following the release of Meta's Llama 2, Nous Research quickly produced Nous-Hermes-Llama2-7b and Nous-Hermes-Llama2-13b, updating the fine-tune to the newer base model.

Hermes 2 (2024)

The Hermes 2 generation represented a significant step forward. Nous Research adopted the ChatML prompt format, which provides a structured system for multi-turn chat dialogue with clear role delineation between system, user, and assistant messages. ChatML enables system prompts that allow users to guide the model's behavior, tone, and persona.

Key Hermes 2 releases included:

• Nous-Hermes-2-Mixtral-8x7B-DPO (January 2024): Fine-tuned on Mistral AI's Mixtral 8x7B mixture-of-experts architecture. This model was trained on over 1,000,000 entries of primarily GPT-4 generated data. It was the first Nous Research model trained with reinforcement learning from human feedback (RLHF) using Direct Preference Optimization (DPO), and the first community model to beat Mixtral Instruct across the majority of popular benchmarks. Both SFT-only and SFT+DPO versions were released.
• Nous-Hermes-2-Yi-34B: Fine-tuned on 01.AI's Yi-34B base model with 200K context length, trained on over 1,000,000 high-quality instruction samples.
• Hermes-2-Pro-Mistral-7B and Hermes-2-Pro-Llama-3-8B: These "Pro" variants introduced dedicated support for function calling and JSON structured output. Hermes 2 Pro scored 90% on a function calling evaluation built in partnership with Fireworks.AI and 84% on structured JSON output evaluation. The models added special tokens (, <tool_call>, <tool_response>) for reliable parsing of agentic interactions.

Nous Research also released the Hermes Function Calling V1 dataset, making public the data mix that gave Hermes 2 Pro its tool use and structured output capabilities. ^[13]

Hermes 3 (August 2024)

Released on August 14, 2024, Hermes 3 was the first full-parameter fine-tune of Meta's Llama 3.1 405B model. The release was made in partnership with Lambda Labs, which provided compute through its 1-Click Cluster infrastructure. ^[4] Hermes 3 was made available in three sizes: 8B, 70B, and 405B parameters.

The training methodology involved three stages:

1. Supervised fine-tuning (SFT) on a large, primarily synthetically generated dataset
2. Reinforcement learning from human feedback (RLHF)
3. Quantization using Neural Magic's FP8 method, reducing VRAM and disk requirements by approximately 50%

For the 405B variant, training was performed on 16 HGX nodes (each containing 8 GPUs) with an effective batch size of 128. The team used the AdamW optimizer with weight decay of 0.01 and a peak learning rate of 3.5 x 10^-6 following a cosine decay schedule after 300 warmup steps over four epochs. ^[3]

Hermes 3 introduced several improvements over Hermes 2, including advanced agentic capabilities, improved roleplaying and internal monologue support, stronger multi-turn conversation coherence, enhanced long-context reasoning, and better overall instruction following. Benchmark results showed performance comparable to or exceeding Meta's official Llama 3.1 Instruct across standard evaluations. At the time of the company's 2025 Series A, Tommy Shaughnessy of Delphi Ventures, an earlier investor, told The Block that Hermes 3 had surpassed 50 million downloads and was powering agents across platforms such as X, Telegram, and gaming environments. ^[6]

DeepHermes 3 (February 2025)

Released in February 2025, DeepHermes 3 Preview was one of the first models to unify both traditional "intuitive" response mode and long chain-of-thought reasoning into a single model. Users could toggle between fast responses and deeper step-by-step reasoning through a system prompt. The model was built on Meta's Llama 3 8B and trained on 1 million non-chain-of-thought responses plus 150,000 chain-of-thought outputs, incorporating 390 million tokens across multiple domains.

Hermes 4 (August 2025)

Hermes 4, released on August 26, 2025, is a family of open-weight, hybrid-reasoning models. ^[5] The family includes models at multiple parameter scales built on different base architectures:

The Hermes 4 technical report states that "Hermes 4 is built upon Llama 3.1 (405B and 70B versions) and Qwen3 14B." ^[16] Hermes 4 introduced a toggleable hybrid reasoning mode via <think>...</think> tags, allowing the model to switch between fast responses and detailed chain-of-thought reasoning. The post-training corpus expanded dramatically compared to Hermes 3, comprising approximately 5 million samples and 19 billion tokens, split into roughly 3.5 million reasoning samples and 1.6 million non-reasoning samples. ^[16] Training was conducted on 192 NVIDIA B200 GPUs using a modified TorchTitan stack with Flex Attention and efficient packing that achieved over 99.9% batch efficiency. ^[16]

Benchmark results for the 405B model (reasoning mode):

RefusalBench is an internal Nous Research benchmark that measures refusal rates across 32 categories using 166 hand-crafted prompts, judged by Claude Sonnet 4. ^[16] The 405B model's RefusalBench score of 57.1 in reasoning mode was the highest among all evaluated models, significantly outperforming GPT-4o (17.67%) and Claude Sonnet 4 (17%). ^[5][16] The release was accompanied by a technical report (arXiv:2508.18255).

Hermes 4.3 (December 2025)

Released in early December 2025, Hermes 4.3 36B is based on ByteDance's Seed-OSS-36B base model. This model is notable for being the first Hermes model trained in a decentralized manner over the internet using the Psyche network. ^[11] Its model card describes the post-training corpus as having grown from about 1 million samples and 1.2 billion tokens for Hermes 3 to roughly 5 million samples and 60 billion tokens. ^[36] Hermes 4.3 supports an extended context length of up to 512K tokens and delivers performance roughly equivalent to Hermes 4 70B at half the parameter count. ^[11]

To test Psyche on a production workload, Nous trained Hermes 4.3 twice: once with its usual centralized recipe (FSDP with AdamW) and once on Psyche (tensor parallelism with DisTrO). The company reported that the Psyche run averaged 144,000 tokens per second across 24 nodes, that the peer-to-peer communication was hidden behind training compute, and that the Psyche-trained model outperformed the centrally trained one on a suite of downstream tasks; it published the centralized version as a research artifact. ^[11] On Nous's RefusalBench, the model card reports that Hermes 4.3 answered 74.6% of questions in non-reasoning mode and 72.3% in reasoning mode, compared with 59.5% for Hermes 4 70B in reasoning mode. ^[36]

Specialist models: Nomos 1 and NousCoder-14B

Alongside the Hermes line, Nous released two task-specific models around the turn of 2026. Nomos 1, published in December 2025 and trained in collaboration with Hillclimb AI, is a specialization of Alibaba's Qwen3-30B-A3B-Thinking-2507 for mathematical problem solving and proof writing in natural language. It is released under the Apache 2.0 license together with an open-source "Nomos Reasoning Harness." According to the model card, Nomos 1 scored 87 out of 120 on the Putnam 2025 problems when run inside that harness, against 24 out of 120 for the base Qwen model under the same conditions. ^[37][52]

NousCoder-14B, released on January 6, 2026, is a competitive-programming model post-trained from Qwen3-14B with reinforcement learning, using an environment built on the company's Atropos framework. Nous reported 67.87% Pass@1 on LiveCodeBench v6 (problems from August 2024 to May 2025), up from 60.79% for the base model, after training on 24,000 verifiable coding problems with 48 NVIDIA B200 GPUs over four days. The write-up compared the DAPO and GSPO objectives and a modified "GSPO+" variant, with DAPO giving the best score at an 81,920-token evaluation context, and Nous released the weights, the RL environment and evaluation harness, and the training logs. ^[38]

Complete Model Timeline

Capybara Series

Alongside the Hermes series, Nous Research developed the Capybara line of models, which took a distinct approach to training data. The Capybara series used a novel data synthesis technique called Amplify-Instruct, which combined several established data synthesis methods including Airoboros, Evol-Instruct, Orca, Vicuna, Know_Logic, Lamini, and FLASK.

The Capybara training dataset was notably compact, containing only 20,000 training examples, roughly 10 times smaller than datasets used for comparable models. Over 60% of the dataset consisted of multi-turn conversations averaging more than 1,000 tokens per example. Seed instructions were drawn from highly regarded datasets like Airoboros, Know Logic, EverythingLM, and GPTeacher, supplemented with new instructions derived from posts on the website LessWrong and in-house multi-turn datasets like Dove.

The most prominent release was Nous-Capybara-34B, fine-tuned on 01.AI's Yi-34B with 200K context length. Smaller versions at 3B and 7B parameters were also released.

Training Methodology

Nous Research has developed and refined a distinctive training approach across its model releases, centered on several key principles.

Synthetic Data Generation

From its earliest models, Nous Research has relied heavily on synthetic data generated by frontier models, particularly GPT-4. The original Hermes model used roughly 300,000 synthetic instructions. By the Hermes 2 generation, training sets grew to over 1,000,000 entries. For Hermes 4, the team employed a graph-based synthetic data generation pipeline that produced approximately 5 million samples totaling 19 billion tokens. ^[16]

The organization has consistently released its training datasets publicly, including the OpenHermes dataset (242,000 GPT-4 generated entries) and the Hermes Function Calling V1 dataset, enabling the broader community to replicate and build upon their work. ^[13]

ChatML Prompt Format

Starting with Hermes 2, Nous Research standardized on the ChatML (Chat Markup Language) prompt format. ChatML uses special tokens to clearly delineate system instructions, user messages, and assistant responses in multi-turn conversations. This structured format enables robust system prompt adherence, clear role boundaries, and reliable parsing of complex interaction patterns including function calling and structured output.

Reinforcement Learning

The Hermes 2 Mixtral release marked Nous Research's first use of RLHF, specifically through Direct Preference Optimization (DPO). Hermes 3 further incorporated RLHF as a dedicated training stage following supervised fine-tuning. By the Hermes 4 generation, the training pipeline included rejection sampling and length control techniques to manage the tendency of reasoning models to produce excessively long outputs.

Neutrally-Aligned Design

A defining characteristic of Hermes models is their "neutrally-aligned" positioning. Unlike many commercial models that implement aggressive content filtering and refusal behavior, Hermes models are trained to follow user instructions with minimal unnecessary refusal. This design philosophy has made Hermes models particularly popular among developers, researchers, and hobbyists who value flexibility and controllability.

Research Contributions

Beyond fine-tuned models, Nous Research has published several notable research papers and technologies.

YaRN (2023)

YaRN (Yet another RoPE extensioN method) is a compute-efficient technique for extending the context window of transformer-based language models that use Rotary Position Embeddings (RoPE). Published by Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole, the paper was first posted as an arXiv preprint in September 2023 and later presented at ICLR 2024. ^[1]

YaRN integrates an "NTK-by-parts" interpolation scheme that selectively scales frequencies to preserve both high-frequency details and local relationships. The method requires only about 0.1% of the original pre-training data for fine-tuning, making it 10x more data-efficient and 2.5x more training-step-efficient than previous context extension methods. ^[1] YaRN achieved state-of-the-art performance in context window extension and has been widely cited in subsequent literature. The technique has been adopted by major model providers including Meta and DeepSeek.

DeMo (2024)

DeMo (Decoupled Momentum Optimization), published in November 2024 by Bowen Peng, Jeffrey Quesnelle, and Diederik P. Kingma (co-creator of the Adam optimizer and OpenAI co-founder), is an algorithm designed to reduce inter-GPU communication during distributed training. ^[2] DeMo uses the Discrete Cosine Transform (DCT) to isolate and share only the most critical components of optimizer momentum, reducing communication overhead by several orders of magnitude while maintaining training performance comparable to AdamW. A revised version posted in February 2026 added Lizhang Chen, Baiyu Su, and Qiang Liu as co-authors and was published at ICLR 2026; it reports that DeMo transmitted up to 85 times less data per GPU than AdamW with distributed data parallelism in experiments on 300-million- and 1-billion-parameter language models, with comparable loss and accuracy. ^[39][40]

DisTrO (2024)

DisTrO (Distributed Training Over-The-Internet) extends the DeMo algorithm into a full system-level stack for large-scale AI training over the internet. Released on August 27, 2024, DisTrO reduces inter-GPU communication bandwidth requirements by 1,000x to 10,000x during pre-training, enabling model training on connections as slow as 100 Mbps download and 10 Mbps upload while maintaining competitive convergence rates. ^[17] In one published test using a Llama 2 architecture, DisTrO reduced per-step communication from 74.4 GB to just 86.8 MB. ^[17]

DisTrO combines low-latency distributed optimizers with a runtime that applies DeMo's momentum-decomposition approach to real-world conditions, supporting compression, selective synchronization, and load balancing across geographically dispersed GPUs. The technology was tested in collaboration with Oracle, Lambda Labs, Northern Data Group, Crusoe Cloud, and the Andromeda Cluster.

Atropos (2025)

Atropos, released in April 2025 under the MIT license, is a framework of reinforcement learning environments for collecting and evaluating language-model trajectories. Nous describes it as the rollout handler of its asynchronous RL pipeline, coordinating generation tasks across potentially thousands of distributed workers through standard inference APIs. ^[41] Later Nous projects built on it: NousCoder-14B's RL environment used Atropos, and in November 2025 the company released tinker-atropos, an integration layer that lets Atropos environments plug into Thinking Machines Lab's Tinker training API. ^[38][42]

Pretraining efficiency research (2026)

In May 2026 Nous published two papers on reducing the cost of pretraining, both by Bowen Peng and Jeffrey Quesnelle with one additional co-author.

• Lighthouse Attention (arXiv:2605.06554, with Subho Ghosh) is a selection-based hierarchical sparse attention method. It pools queries, keys, and values symmetrically across a multi-resolution pyramid, picks a small dense sub-sequence using per-head vector norms, and runs standard FlashAttention on that sub-sequence, so no custom sparse kernel is needed. A short phase of ordinary attention at the end of training converts the model back into a dense-attention model. Nous reported a forward and backward pass about 17 times faster than standard attention at 512K context on a single B200 GPU, and a 1.4 to 1.7 times end-to-end pretraining speedup at 98K context. ^[43][44]
• Token Superposition Training (arXiv:2605.06546, with Théo Gigant) has the model read and predict averaged "bags" of contiguous tokens during the first 20 to 40 percent of training, then return to ordinary next-token prediction, so the final model is identical in form to a conventionally trained one. Nous reported a 2 to 3 times wall-clock speedup at matched FLOPs, tested on dense models from 270 million to 3 billion parameters and on a 10B-A1B mixture-of-experts model. ^[45][46] Quesnelle's homepage lists the paper among NeurIPS 2026 publications. ^[33]

A January 2026 engineering post described the company's training stack as a heavily modified fork of TorchTitan and documented its integration of DeepSeek's DeepEP expert-parallelism library for mixture-of-experts pretraining on a B200 cluster. ^[47]

Contrastive neuron attribution (2026)

In May 2026 Nous described contrastive neuron attribution (CNA), an interpretability and steering method that uses pairs of contrasting prompts to locate sparse circuits of MLP neurons associated with a behavior, building on work by Transluce showing that such circuits sit in the MLP neuron basis. Using refusal as its case study, Nous reported that as few as eight contrastive prompt pairs could identify a circuit for a behavior and that steering through it left MMLU performance unchanged, and it presented the method as an alternative to residual-stream control vectors and sparse autoencoders. ^[48]

What is the Psyche Network?

Psyche is Nous Research's decentralized AI training orchestration platform, built on the Solana blockchain. The network applies DisTrO technology to enable collaborative model training across underutilized hardware distributed around the world. Nous Research describes Psyche as "an open infrastructure that democratizes AI development by decentralizing training across underutilized hardware." ^[14]

Solana serves as the coordination hub for Psyche. Smart contracts on the blockchain store training metadata, participant lists, and random task assignments, providing transparency, tamper-proofing, and censorship resistance. The design creates a fault-tolerant system where individual nodes can join or leave training runs without disrupting the overall process.

In December 2024, Nous Research ran a successful test of the Psyche network, training a 15 billion parameter language model through 11,000 training steps using hardware distributed across multiple locations. ^[10] In 2025, the network's first major production run pretrained Consilience, a 40 billion parameter model, from scratch over the internet across 20 trillion tokens, using a DeepSeek v3-style Multi-head Latent Attention (MLA) architecture (a dense variant without mixture-of-experts routers). ^[18] The training mix combined FineWeb (14T tokens), FineWeb-2 (4T tokens), and The Stack v2 (about 0.2T tokens upsampled to 1T), with model checkpoints automatically updated every 500 training steps. ^[18] Nous Research framed Consilience as an attempt to "make a true 'base' model, one representative of the entirety of the creative output of humanity, and not merely trying to win the benchmaxxing game." ^[18] According to research group Epoch AI, Consilience 40B ranks among the largest decentralized pretraining runs to date, alongside Prime Intellect's INTELLECT-1 and Pluralis Research's Protocol Model. ^[19]

Nous announced the Psyche testnet on April 25, 2025, the same day as its Series A. ^[6] On September 29, 2025, the Psyche team said the network would begin training several models in parallel, moving "from provability to performance." Code changes, including a full trainer abstraction, let Psyche run supervised fine-tuning as well as pretraining. The plan included a series of pretraining data-mix ablations drawing on open datasets such as FineWeb-Edu, FineWeb2-HQ, FinePDFs, DCLM-Baseline, Stack-Edu, and Nemotron-CC-Math, and a Hermes 4 post-training run on ByteDance's Seed-OSS-36B. The same post called the Consilience 40B run the largest distributed pretraining run to that date. ^[49] That Seed-OSS run produced Hermes 4.3 36B, released in December 2025, the first Hermes model post-trained entirely on the Psyche network. ^[11]

As of early 2025, Nous Research had not yet launched a dedicated NOUS token but was evaluating whether to reward participants with a proprietary token or with Solana's native cryptocurrency. ^[8]

Forge Reasoning API

The Forge Reasoning API, introduced in November 2024, is a configurable step-based planning engine that combines multiple open and closed LLMs with Nous Research's reasoning techniques. The system employs Monte Carlo Tree Search (MCTS), Chain-of-Code (CoC) reasoning, and a Mixture of Agents (MoA) architecture to enable flexible, multi-step problem solving.

Forge is distinct from the fine-tuned model releases; it functions as a reasoning framework that orchestrates multiple models to tackle complex tasks that require planning and iterative refinement.

Nous Portal and Hermes Agent

Nous Portal

Nous Research launched its first-party inference service on March 12, 2025, an OpenAI-compatible completions and chat completions API served through a new account hub called the Nous Portal at portal.nousresearch.com. The initial release offered two models, Hermes 3 Llama 70B and DeepHermes 3 8B Preview, behind a first-come, first-served waitlist, and every account started with $5.00 of free credits. ^[20] The company said it would begin with "the cheapest inference we can offer at cost-basis for Hermes models" and adjust the offering as the Psyche network's infrastructure grew. ^[20]

By mid-2026 the Portal had expanded from a Hermes inference API into a paid subscription gateway built around Hermes Agent. Nous Research's documentation describes the Portal as the company's "unified subscription gateway" and the recommended way to run Hermes Agent: a single login provides access to more than 300 models from multiple providers, including Claude, GPT, Gemini, and DeepSeek models, plus a "Tool Gateway" that routes web search, image generation, text-to-speech, browser automation, and terminal sandbox services through the same subscription. ^[21] As of July 2026, hosted paid tiers ranged from $20 to $200 per month. ^[23] In October 2026 the Hermes Agent website listed Plus ($20), Super ($100), and Ultra ($200) monthly plans, each including monthly credits worth 10 percent more than the price, alongside a free tier limited to free models. ^[50] The Hermes Agent documentation advises against using Nous's own Hermes 4 models inside the agent, saying they are tuned for chat and reasoning rather than the rapid tool-calling loop the agent relies on. ^[21]

Hermes Agent

Hermes Agent is Nous Research's open-source AI agent, released under the MIT License in February 2026 and developed in public on GitHub, where versioned releases began in March 2026. ^[22][32] Nous Research's release log dates the launch to February 25, 2026. ^[52] The agent is built around a persistent learning loop: it creates reusable skills from its own experience, searches its past conversations, and maintains memory of the user across sessions, and it can be used from a terminal or through messaging platforms including Telegram, Discord, Slack, WhatsApp, and Signal. ^[22] The project became one of the most popular repositories on GitHub during 2026: TechCrunch reported roughly 214,000 stars and nearly 40,000 forks in July 2026, and the count passed 225,000 stars in early August 2026. ^[22][23]

By October 27, 2026, the repository showed about 257,000 stars and 56,000 forks. ^[53] In its October 2026 fundraising note, Nous said Hermes had been cloned more than 24 million times and, by the company's internal estimates, drove about 2.5% of global token usage; The Next Web reported that The Wall Street Journal counted more than 22 million downloads since February. ^[32][35] The Block, summarizing TechCrunch's July reporting, described Hermes as a direct competitor to OpenClaw and noted that Nous released it weeks after OpenClaw's agent went viral. ^[56]

The project ships frequent tagged releases. Selected milestones from its GitHub release notes: ^[51]

In September 2026 Teknium described using Hermes Agent to refactor its own codebase. On September 2 he asked his everyday Hermes agent to simplify the repository; over about nineteen active hours it dispatched 1,393 subagents, with up to 218 running at once, while Claude Fable 5.1 handled inference. After two rounds of community review the pull request was merged on September 4, cutting non-test Python from 1,063,826 to 698,363 lines (34.4%). He put the model cost at about $19,300 for the main run, or roughly $25,000 including follow-up sessions, and wrote that reviewers caught real regressions that the existing tests had missed. ^[54]

Hermes for Businesses

The October 2026 funding was earmarked for Hermes for Businesses. ^[31][32] Nous's Hermes Business page describes a shared deployment for a team: one central balance with per-member spending caps, hosted agents for every member, and skills that any member builds published to a team library for everyone's agent, all in an isolated tenant per team on Nous infrastructure. A separate Enterprise tier runs on infrastructure the customer controls, with SSO, service-level agreements, and assisted onboarding. ^[55]

Partnership with Novita AI

Since May 2026 Nous Research has partnered with Novita AI, a GPU cloud and inference provider whose platform offers more than 200 models through a serverless API alongside GPU instances and agent sandboxes, to run discounted third-party model offerings on the Nous Portal. ^[24][30] On May 14, 2026, Nous Research made DeepSeek's V4 Flash model free through the Portal for a limited time, crediting Novita, and Novita promoted the model as "powered by Novita AI" for use in Hermes Agent. ^[24][25] On August 2, 2026, Nous Research announced that DeepSeek V4-Flash 0731, the official build DeepSeek had released two days earlier, was available on the Portal at a 90 percent discount for seven days, again in partnership with Novita. ^[26]

The August 2026 announcement claimed that at the discounted price the model was "over 1000x cheaper than Fable 5 on comparable tasks while still beating it on Terminal-Bench 2.1," a comparison with Anthropic's flagship Claude Fable 5 model. ^[26] These are Nous Research's own marketing claims, and the benchmark comparison is not supported by the public leaderboard: as of August 5, 2026, the official Terminal-Bench 2.1 leaderboard at tbench.ai listed a Claude Code and Fable 5 configuration in first place at 83.8% and included no DeepSeek entry, while DeepSeek's self-reported Terminal-Bench 2.1 score for V4-Flash-0731, measured with an unreleased in-house harness, was 82.7, and an independent run by Artificial Analysis measured 79%. ^[27][28][29]

Community and Ecosystem

Nous Research maintains an active Discord community with about 147,000 members as of October 2026, ^[57] where researchers, developers, and enthusiasts discuss open-source AI development. The community has been central to the organization's identity since its founding as a volunteer-driven project.

All Hermes models and most training datasets are released publicly through the NousResearch organization on Hugging Face. ^[9] Models are also available through inference providers such as OpenRouter, Lambda Labs, Fireworks.AI, and Together AI, and are compatible with local inference tools like Ollama and LM Studio through GGUF quantized versions.

In January 2024 Nous Research launched a fine-tuning subnet on the Bittensor network, where miners are rewarded for fine-tuning language models using continuously generated synthetic data; its documentation describes it as the first continuous, incentivized fine-tuning benchmark. ^[58]

How does Nous Research fit into the open-source AI landscape?

Nous Research occupies a distinct niche in the open-source AI ecosystem. While organizations like Meta, Mistral AI, and 01.AI release base models and official instruction-tuned variants, Nous Research specializes in community-driven post-training that often pushes these base models beyond their official fine-tunes in specific capabilities. The Hermes series has frequently been among the top-performing models on the Hugging Face Open LLM Leaderboard and other community benchmarks.

The organization's approach differs from other fine-tuning groups in several ways. Its emphasis on synthetic data pipelines, ChatML standardization, function calling support, and neutrally-aligned behavior has created a recognizable "Hermes" style that many users prefer for agentic applications, creative writing, and unconstrained research tasks.

With its Series A funding and Psyche network development, Nous Research has also positioned itself at the intersection of AI and decentralized AI computing, pursuing a vision where model training is not concentrated in a handful of large data centers but distributed across globally contributed hardware. In 2026 Hermes Agent became the company's best-known product, and the Series B was raised to build a business version of it. ^[31][32]

Key People

See also

• Hermes Agent
• Hermes 4
• YaRN
• OpenClaw
• Prime Intellect
• List of neolabs
• Neolab

References

1. ^1 ^2Peng, B., Quesnelle, J., Fan, H., & Shippole, E. (2023). "YaRN: Efficient Context Window Extension of Large Language Models." arXiv:2309.00071. Presented at ICLR 2024.
2. ^Peng, B., Quesnelle, J., & Kingma, D. P. (2024). "DeMo: Decoupled Momentum Optimization." arXiv:2411.19870v1, November 29, 2024. arxiv.org/...2411.19870v1
3. ^Teknium et al. (2024). "Hermes 3 Technical Report." arXiv:2408.11857. Nous Research.
4. ^"Unveiling Hermes 3: The First Full-Parameter Fine-Tuned Llama 3.1 405B Model is on Lambda's Cloud." Lambda AI Blog, August 15, 2024.
5. ^1 ^2"Nous Research Team Releases Hermes 4: A Family of Open-Weight AI Models with Hybrid Reasoning." MarkTechPost, August 27, 2025.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Paradigm leads $50 million Series A round for decentralized AI project Nous Research." The Block, April 25, 2025. theblock.co/...entralized-ai-project-nous-research
7. ^"Nous Research Raises $65M in Funding." FinSMEs, April 2025.
8. ^1 ^2 ^3 ^4 ^5"Exclusive: Crypto VC giant Paradigm makes $50 million bet on decentralized AI startup Nous Research at $1 billion token valuation." Fortune Crypto, April 25, 2025. fortune.com/...-capital-deepseek-openai-blockchain
9. ^NousResearch Organization, Hugging Face. huggingface.co/NousResearch
10. ^"Nous Research is training an AI model using machines distributed across the internet." VentureBeat, December 2, 2024.
11. ^1 ^2 ^3 ^4"Introducing Hermes 4.3: Local Intelligence Globally Trained." Nous Research Blog, December 2025. nousresearch.com/introducing-hermes-4-3
12. "Data synthesis for SOTA LLMs with Karan Malhotra, researcher at Nous Research." Practical AI Podcast, Episode 255, The Changelog.
13. ^1 ^2NousResearch/Hermes-Function-Calling GitHub Repository. github.com/...Hermes-Function-Calling
14. ^"Democratizing AI: The Psyche Network Architecture." Nous Research Blog, 2025. nousresearch.com/nous-psyche
15. ^"Nous Research raised $5.2M Seed Funding on Jan 10, 2024." CypherHunter / Nous Research announcement, January 2024.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7Teknium, R., Jin, R., et al. (2025). "Hermes 4 Technical Report." arXiv:2508.18255. Nous Research, August 26, 2025.
17. ^1 ^2"This could change everything! Nous Research unveils new tool to train powerful AI models with 10,000x efficiency." VentureBeat, August 27, 2024; NousResearch/DisTrO GitHub Repository, github.com/...DisTrO.
18. ^1 ^2 ^3PsycheFoundation/consilience-40b, Hugging Face model card. huggingface.co/...consilience-40b-CqX3FUm4
19. ^"How far can decentralized training over the internet scale?" Epoch AI, Gradient Updates, 2025. epoch.ai/...lized-training-over-the-internet-scale
20. ^1 ^2 ^3Kainan. "Announcing the Nous Portal." Nous Research Forums, March 12, 2025. forum.nousresearch.com/...62
21. ^1 ^2"Nous Portal." Hermes Agent documentation, Nous Research. Accessed August 5 and October 27, 2026. hermes-agent.nousresearch.com/...nous-portal
22. ^1 ^2 ^3 ^4NousResearch/hermes-agent GitHub repository. Accessed August 5, 2026. github.com/...hermes-agent
23. ^1 ^2 ^3 ^4 ^5"Hermes agent maker Nous Research in talks for new funding at $1.5B valuation." TechCrunch, July 13, 2026. techcrunch.com/...or-new-funding-at-1-5b-valuation
24. ^1 ^2Nous Research (@NousResearch) on X, May 14, 2026. x.com/...2055051105432752509
25. ^Novita AI (@novita_labs) on X, May 19, 2026. x.com/...2056776161074798629
26. ^1 ^2Nous Research (@NousResearch) on X, August 2, 2026. x.com/...2083953441571742191
27. ^Terminal-Bench 2.1 leaderboard, tbench.ai. Accessed August 5, 2026. tbench.ai/...2.1
28. ^"Change Log" (entry dated July 31, 2026), DeepSeek API Docs. api-docs.deepseek.com/updates
29. ^Artificial Analysis (@ArtificialAnlys) on X, July 31, 2026. x.com/...2083123180869496865
30. ^Novita AI website. Accessed August 5, 2026. novita.ai
31. ^1 ^2 ^3 ^4 ^5 ^6 ^7Temkin, Marina. "Nous Research confirms it hit $1.5B valuation, launches AI agents for business users." TechCrunch, October 7, 2026. techcrunch.com/...hes-ai-agents-for-business-users
32. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Rolnick, Dillon. "A Note on Our Fundraise." Nous Research Blog, October 7, 2026. nousresearch.com/a-note-on-our-fundraise
33. ^1 ^2 ^3 ^4Quesnelle, Jeffrey. "Homepage of Jeffrey Quesnelle." Accessed October 27, 2026. jeffq.com
34. ^Dillon Rolnick (@dillonrolnick) on X, profile description. Accessed October 27, 2026. x.com/dillonrolnick
35. ^1 ^2Dina, Cristian. "Nous Research raises $90M to take its Hermes AI agent to businesses." The Next Web, October 7, 2026. thenextweb.com/...earch-90m-1-5bn-hermes-agent-wsj
36. ^1 ^2NousResearch/Hermes-4.3-36B, Hugging Face model card. huggingface.co/...Hermes-4.3-36B
37. ^NousResearch/nomos-1, Hugging Face model card. huggingface.co/...nomos-1
38. ^1 ^2Li, Joe. "NousCoder-14B: A Competitive Olympiad Programming Model." Nous Research Blog, January 6, 2026. nousresearch.com/...ive-olympiad-programming-model
39. ^Peng, B., Chen, L., Su, B., Quesnelle, J., Kingma, D. P., & Liu, Q. "DeMo: Decoupled Momentum Optimization." arXiv:2411.19870v2, February 6, 2026. arxiv.org/...2411.19870
40. ^"DeMo: Decoupled Momentum Optimization." International Conference on Learning Representations 2026 (ICLR 2026), proceedings. proceedings.iclr.cc/...5da3657-Abstract-Conference
41. ^"Introducing Atropos." Nous Research Blog, April 29, 2025. nousresearch.com/introducing-atropos; NousResearch/atropos GitHub repository, github.com/...atropos
42. ^"tinker-atropos: An Integration Layer for the Tinker API and Atropos RL Framework." Nous Research Blog, November 2025. nousresearch.com/tinker-atropos-blog
43. ^Peng, B., Ghosh, S., & Quesnelle, J. "Long Context Pre-Training with Lighthouse Attention." arXiv:2605.06554, May 7, 2026. arxiv.org/...2605.06554
44. ^"Lighthouse Attention." Nous Research Blog, May 12, 2026. nousresearch.com/lighthouse-attention
45. ^Peng, B., Gigant, T., & Quesnelle, J. "Efficient Pre-Training with Token Superposition." arXiv:2605.06546, May 7, 2026. arxiv.org/...2605.06546
46. ^"Efficient pretraining with token superposition." Nous Research Blog, May 13, 2026. nousresearch.com/token-superposition
47. ^Nguyen, Phuc. "Field Notes on Scaling MoE Expert Parallelism with DeepEP." Nous Research Blog, January 20, 2026. nousresearch.com/moe-scaling-field-notes
48. ^Nightwing. "Model Neuroscience: Dissecting Behavioral Change With Targeted Contrastive Neuron Attribution." Nous Research Blog, May 14, 2026. nousresearch.com/neuron-steering
49. ^Psyche Team. "The Next Phase of Psyche." Nous Research Blog, September 29, 2025. nousresearch.com/the-next-phase-of-psyche
50. ^Hermes Agent website, Nous Research (pricing section). Accessed October 27, 2026. hermes-agent.nousresearch.com
51. ^NousResearch/hermes-agent releases, GitHub (release notes for v0.2.0, v0.10.0, v0.16.0, v0.21.0, and v0.22.0). Accessed October 27, 2026. github.com/...releases
52. ^1 ^2"Releases." Nous Research. Accessed October 27, 2026. nousresearch.com/releases
53. ^NousResearch/hermes-agent repository record (star and fork counts), GitHub REST API. Accessed October 27, 2026. api.github.com/...hermes-agent
54. ^Teknium. "Refactoring Hermes with 1,393 agents." Nous Research Blog, September 15, 2026. nousresearch.com/refactoring-hermes-with-1393-agents
55. ^"Hermes Business." Nous Portal. Accessed October 27, 2026. portal.nousresearch.com/business
56. ^"Decentralized AI project Nous Research in talks to raise $75M at $1.5B valuation: report." The Block, July 14, 2026. theblock.co/...408237
57. ^Nous Research Discord server invite (approximate member count). Accessed October 27, 2026. discord.gg/nousresearch
58. ^NousResearch/finetuning-subnet GitHub repository (README). github.com/...finetuning-subnet
59. ^"Careers." Nous Research. Accessed October 27, 2026. nousresearch.com/careers
60. ^"Terms of Service." Nous Research. Accessed October 27, 2026. nousresearch.com/terms
61. ^Teknium (@Teknium) on X, profile description. Accessed October 27, 2026. x.com/Teknium
62. ^Nous Research (@NousResearch). Announcement of the closing of its $5.2 million seed financing round. X, January 9, 2024. x.com/...1744865872563618128