Nous Research and Psyche, the open-source, decentralized AI revolution

Note: As an amusing aside, this article was partly written using Nous Research's Hermes model, which we used through the Nous Chat interface.

---

What is Nous Research?

Nous Research is a collective of artificial intelligence (AI) researchers and developers formed around a shared vision: prioritizing the development of open-source models and promoting distributed, decentralized systems to make these technologies more accessible.

Currently, the majority of AI models and software are proprietary, meaning the code belongs to the company that developed it. While there are initiatives to promote open source, notably supported by Meta, this momentum could stall, leaving the entire open-source community dependent on fully centralized solutions.

Nous Research aims to conduct fundamental research on training, fine-tuning, and inference of AI models to make these technologies more accessible without reliance on centralized corporations.

Nous Research is not structured as a traditional company. It is a group of researchers, developers, and AI enthusiasts who initially connected via Discord servers or Reddit before uniting to undertake more ambitious initiatives. The project has raised $5.2 million and recently received a grant from the renowned a16z fund as part of their Open Source AI Grant Program.

Nous Research is behind several significant products and innovations (discussed further below), including:

• Hermes 3: An optimized language model, more efficient and less biased than its competitors.
• DisTrO: A framework for distributed AI model training via GPUs located worldwide.
• Forge API: An API combining multiple models for richer and more diverse results.
• Nous Chat: A user interface for a simplified experience with Hermes 3.
• Nous Psyche: The latest initiative, aimed at democratizing collaborative AI training on Solana.

---

Nous Psyche: Collaborative AI Training on Solana

On January 27, 2025, Nous Research announced the launch of Nous Psyche, marking a new step in its mission to democratize AI model training and inference. Nous Psyche is a decentralized platform, built on distributed infrastructure, designed to enable anyone to participate in training cutting-edge AI models.

A Decentralized Training Network

The first phase of Psyche leverages Nous DisTrO, a distributed training framework that allows various types of GPUs to collaboratively contribute to AI model training. The Solana blockchain plays a central role in coordinating participant contributions, ensuring the integrity and traceability of training tasks.

One of Psyche’s key aspects is its fault tolerance and flexibility. During initial tests, Nous Research demonstrated that new compute nodes could be dynamically added during training and that the system could absorb individual node failures without disrupting the overall process.

Initially, the Psyche network is semi-permissioned, meaning compute providers must be approved by Nous Research. The long-term goal is to transition to a fully permissionless network where anyone can contribute to model training.

Decentralized Inference and Reinforcement Learning

While the first phase of Psyche focuses primarily on collaborative model training, the second phase will introduce accessible inference infrastructure and an advanced reinforcement learning (Reinforcement Learning) approach.

A major challenge in developing sophisticated AI models is their ability to autonomously improve reasoning. Psyche aims to integrate distributed reinforcement learning methodologies, where inference is an integral part of the training process. This architecture fundamentally differs from conventional approaches, where training and inference are treated separately.

With this approach, Nous Research plans to pre-train, fine-tune, and serve its future language models directly on Psyche. This system will establish a continuous improvement process for models while ensuring instant distribution of AI advancements.

Optimization with Nous DisTrO

Psyche is built on Nous DisTrO, one of Nous Research’s most advanced developments. This technology significantly reduces bandwidth requirements and optimizes GPU communication. In a large-scale test, DisTrO enabled:

• A bandwidth reduction factor of 1,000x to 10,000x compared to conventional solutions.
• The elimination of costly hardware interconnections while maintaining high performance.
• System robustness, with more than 11,000 successful training steps, and the dynamic addition of new nodes boosting training rates without compromising stability.

DisTrO’s development was led by some of the most influential researchers in the field, including Diederik P. Kingma, co-founder of OpenAI and co-author of the Adam optimization algorithm, widely used in modern AI model training.

The Role of Solana and Psyche’s Economic Model

One of Nous Research’s most strategic decisions was to implement Psyche on Solana, a network known for its low transaction costs and scalability. According to a post by @0xElonMoney, the Psyche network operates with an incentive system that rewards contributors based on their participation:

• Rewards for model training: Participants who provide compute power receive tokens in exchange for their contributions.
• Rewards for validation and monitoring: To ensure network integrity, a validation mechanism allows participants to verify the quality of others’ contributions.
• Penalties for dishonest behavior: A penalty system reduces rewards for nodes that fail to meet protocol requirements (e.g., submitting incorrect data).

While no official token has been announced, it is highly likely that Nous Research will implement one to incentivize contributors and maintain Psyche’s ecosystem. Furthermore, in the documentation for Venice’s VVV token launch, an allocation is mentioned for “the launch of Psyche’s token.”

A Turning Point for Decentralized AI Development?

With Psyche, Nous Research is not merely improving existing models—it aims to fundamentally reshape how AI is trained and distributed. The project introduces a decentralized infrastructure that could:

• Reduce dependence on large data centers, making AI model training more accessible.
• Encourage a collaborative approach, where researchers, companies, and individuals contribute to model improvement.
• Ensure resilience against censorship and regulatory restrictions, leveraging a distributed and autonomous network.

Nous Research plans to integrate Psyche across its entire ecosystem, including hosting future Hermes models directly on the network. This includes Hermes 3 and its subsequent versions, which will be fine-tuned and improved using Psyche’s distributed infrastructure.

Psyche’s future may also involve integrating multimodal models, capable of processing text, images, and videos, as well as applications in gaming, simulation, and autonomous AI agents.

By combining open-source AI, distributed infrastructure, and token-based economic incentives on Solana, Nous Psyche could become one of the most ambitious projects in the decentralized AI ecosystem.

Other Nous Research Products

→ For more information about Nous Research, check out the full video from our contributor, Victor (@victor_crypto_).

Hermes 3

One of Nous Research’s most notable developments is Hermes 3, a family of language models optimized from Llama 3 (the model developed by Meta). Through advanced fine-tuning, Nous Research has enhanced several key aspects of the model, including reasoning capabilities, adherence to instructions, and long-context retention.

Unlike dominant models developed by centralized companies, Hermes 3 was designed to be less biased and more neutral in its responses. Nous Research has highlighted this feature by publishing comparisons with other LLMs (such as Claude and GPT), demonstrating that Hermes 3 does not artificially filter certain topics and respects user requests without ideological distortion.

Its success has not been limited to research. Hermes 3 is now one of the most downloaded open-source models (over 50 million downloads), powering a large portion of AI agents on X (formerly Twitter), Discord, Telegram, and other platforms.

The training of Hermes 3 was carried out in collaboration with Lambda Labs, and the model is accessible via Hugging Face for researchers and developers looking to integrate it into their applications.

---

DisTrO

Nous Research has also developed DisTrO (Distributed Training Over-the-Internet), a collaborative training framework that enables pre-training and fine-tuning of AI models in a decentralized manner, using GPUs distributed worldwide.

The main challenge in training AI models is that it requires enormous computing resources, often centralized in data centers owned by a few large corporations. DisTrO offers a more open alternative, allowing AI models to be trained on distributed infrastructures without requiring expensive high-bandwidth interconnections.

DisTrO’s performance was tested on a large scale by training a 15-billion-parameter model, utilizing GPUs provided by Oracle, Northern Data, Lambda Labs, Crusoe Cloud, and Andromeda Cluster. The experiment demonstrated several major breakthroughs:

• Compression of bandwidth usage by a factor of 1,000x to 10,000x compared to conventional methods, making training accessible with a standard internet connection.
• Fault tolerance: The system can adapt to interruptions without disrupting the overall training process.
• Scalability: New computing nodes can be dynamically added to accelerate training without any stability loss.

The algorithm behind DisTrO, called DeMo (Decoupled Momentum Optimization), was developed in collaboration with Diederik P. Kingma, co-founder of OpenAI and co-author of the Adam optimization algorithm, one of the most widely used in deep learning.

Note: DisTrO serves as the technological foundation for Nous Psyche, but its use extends far beyond this project. It represents a major advancement in collaborative machine learning and open science in AI.

---

Forge API

Another key project from Nous Research is Forge API, an API that orchestrates multiple AI models to enhance the quality of generated responses. Instead of relying on a single LLM, Forge API employs a Mixture of Models (MoM) approach, where several specialized models work together to provide more precise and nuanced results.

Specifically, Forge API enables:

• Combining models like Hermes 3, GPT, Claude, or Gemini to enhance overall performance.
• Defining AI agent chains, where multiple models interact to solve complex tasks.
• Optimizing responses by dynamically selecting the most suitable model based on the context.

This multi-model approach significantly improves the robustness and efficiency of AI systems, particularly in applications such as advanced research, information synthesis, and decision analysis.

---

Nous Chat

Nous Research recently launched Nous Chat, an interface designed for easy interaction with Hermes 3, similar to ChatGPT or HuggingChat.

Previously, users had to download and run the model on their own machines or access it via partner platforms. With Nous Chat, Hermes 3 is now easily accessible through a streamlined web interface.

Nous Chat stands out with several features:

• A minimalist and fluid interface, inspired by early computer terminals.
• Direct access to Hermes 3-70B, with minimal response filtering.
• Predefined prompt categories, allowing users to explore various AI applications (analysis, creative writing, problem-solving, etc.).

Unlike solutions from OpenAI or Anthropic, Nous Chat does not collect user data and remains committed to the privacy-first, open-source philosophy championed by Nous Research.

Simulators

Nous Research is also exploring interactive and immersive AI concepts, particularly through Simulators.

These AI systems aim to replicate dynamic and autonomous environments, where models can learn by directly interacting with their surroundings. Some ongoing experiments include:

• TEE, an AI agent on X (Twitter) capable of proving its true autonomy.
• God & Satan, two highly followed X accounts that were previously managed by individuals. These are now being trained and fine-tuned to become fully autonomous AI-driven entities, generating and interacting with content.
• The "World Simulator", a project designed to simulate complex AI-driven scenarios, including social and economic dynamics.

---

Conclusion

Nous Research has established itself as one of the most ambitious collectives in decentralized artificial intelligence. While tech giants like OpenAI, Anthropic, and Google advance with billions in funding and centralized infrastructure, Nous Research is charting its own path, driven by a radical open-source vision and the sharing of computing resources.

With Nous Psyche, the team aims to make AI model training as accessible as crypto mining in its early days. If the project reaches its full potential, it could reshape the AI landscape, giving researchers, developers, and everyday users an active role in model development.

But one question remains: Can this approach compete with the optimization power of Big Tech? The future will tell if decentralization and open collaboration can offer a viable alternative to supercomputer clusters and closed-source models.