Nous Research officially announced the launch of Psyche, a decentralized AI training network based on the Solana blockchain, aiming to enable global users to participate in the development of cutting-edge AI models using idle computing resources. This platform uses DisTrO technology to significantly reduce bandwidth requirements and successfully initiated the largest distributed training ever conducted over the internet, with the goal of training a 4 billion-parameter AI model, all data and checkpoints being fully open-source. AIbase delves into the core technologies and vision of Psyche, revealing its profound impact on democratizing AI.
The Psyche Network: A Breakthrough Platform for Universal AI Development
The Psyche network is centered around decentralization, breaking away from the traditional reliance on centralized data centers for AI training. The platform coordinates global distributed GPUs through the Solana blockchain, allowing anyone to contribute idle computing resources to participate in model training. AIbase learned that Psyche's first major task is to pre-train a 4 billion-parameter Consilience model using the Multi-Head Potential Attention (MLA) architecture, based on a super-large-scale dataset of 20 trillion tokens (including FineWeb14T, FineWeb-24T, and The Stack v21T). This marks the largest distributed AI training ever conducted on the internet, signifying a milestone in decentralized AI development.
Nous Research emphasizes that the Consilience model is designed to balance efficiency and accessibility. Its 4 billion parameter scale allows it to be trained on a single H/DGX server or run inference on consumer-grade RTX3090 GPUs, while also supporting long context processing, laying the foundation for advanced reasoning and creative applications. AIbase believes this scale precisely balances performance with popularization goals, providing powerful tools for small-to-medium teams and individual developers.
Core Technology: The Synergy of DisTrO and Solana
The success of the Psyche network relies heavily on Nous Research’s DisTrO (Distributed Training Over-the-Internet) technology. This technology optimizes communication between GPUs by compressing inter-GPU data volume by 1,000 to 10,000 times, similar to the frequency domain transformation principle of JPEG image compression. AIbase analyzes that DisTrO allows each node to train independently without step-by-step synchronization, significantly reducing network latency and bandwidth requirements, enabling heterogeneous hardware (such as RTX4090, A100, H100) to work collaboratively.
Solana blockchain plays the role of a coordination hub in Psyche. Smart contracts store training metadata, participant lists, and random task assignments, ensuring transparency, tamper-proofing, and censorship-resistance in the training process. AIbase notes that Solana's high throughput and low transaction costs make it an ideal platform for handling distributed training micro-transactions, further enhancing Psyche's global scalability.
Open Source and Community-Driven: Democratizing AI Development
Psyche is committed to promoting AI democratization through complete openness. All training data, checkpoints, and results will be publicly shared, hosted on Hugging Face and GitHub, available for free access by global developers. AIbase learned that Psyche's code has been open-sourced, allowing developers to participate in optimization via Rust-based systems and P2P networks or build custom applications based on open documentation.
Nous Research plans to further expand Psyche's functionalities, including support for model fine-tuning and community-driven data inputs, allowing small teams and even individuals to customize their own AI models. For example, communities can upload domain-specific data to train models tailored to education, healthcare, or creative fields. AIbase predicts that this community-driven model will stimulate global innovation, challenging the monopolistic positions of centralized AI giants like OpenAI and DeepSeek.
Industry Impact: The Rise of Decentralized AI
The launch of Psyche marks the rapid development phase of decentralized AI. AIbase analyzes that compared to traditional cloud-based AI, Psyche's distributed architecture not only reduces training costs but also transforms power contributions from "donation" to "transaction" through blockchain incentive mechanisms, attracting widespread participation. In December 2024, Psyche successfully completed a test training of a 15 billion-parameter model spanning 11,000 steps, verifying its stability across the global network.
In April 2025, Nous Research secured a $50 million Series A financing led by Paradigm, valuing the company at $1 billion, further accelerating Psyche's development. AIbase noticed that within 44 minutes of launching the testnet, the project raised $500,000 worth of GPU power through donations, reflecting the high enthusiasm of the community for Psyche.
A Turning Point for AI Democratization
As an authoritative media in the AI field, AIbase believes that the launch of Psyche is not only a milestone for Nous Research but also a turning point for decentralized AI. Its innovative architecture combining DisTrO and Solana provides global developers with opportunities to participate in AI development, breaking down technical and resource barriers. Of particular note is the potential impact of Psyche on China's AI community; native models like Qwen3 may achieve more efficient localization fine-tuning through its open-source framework.