Psyche: Decentralized AI Training Platform Ushers in a New Era of Mass Development

Nous Research recently announced the launch of Psyche, a decentralized AI training network built on the Solana blockchain. The goal of this platform is to enable global users to contribute their idle computing resources to the development of advanced AI models. By utilizing DisTrO (Distributed Training Over-the-Internet) technology, Nous Research has significantly reduced bandwidth requirements, making it possible to train a 40-billion-parameter AI model in one of the largest internet-distributed training efforts to date. This pioneering network aims to open up AI development to a broader audience, democratizing access and participation in the field. A Breakthrough Platform for Global AI Development Psyche's core strength lies in its decentralized approach, which eliminates the need for centralized data centers typically associated with AI training. The platform leverages the Solana blockchain to coordinate a global network of GPUs, allowing anyone to contribute their idle computational capacity. According to AIbase, Psyche's first major task is to pre-train a 40-billion-parameter model called Consilience. This model uses a multi-head latent attention (MLA) architecture and is trained on an enormous dataset of 20 trillion tokens, including FineWeb14T, FineWeb24T, and The Stack v21T. Consilience is designed to be both efficient and accessible. Its 40-billion-parameter size allows it to be trained on a single H/DGX server or run inference on consumer-grade RTX 3090 GPUs. Additionally, the model supports long context processing, which is crucial for advanced reasoning and creative applications. AIbase notes that this design strikes a balance between performance and inclusivity, providing powerful tools for small teams and individual developers. Technological Foundation: DisTrO and Solana Collaboration The success of the Psyche network hinges on Nous Research's innovative DisTrO technology. DisTrO optimizes the communication between GPUs by compressing the data transmitted during training, similar to how JPEG images are compressed in the frequency domain. This reduces the GPU-to-GPU communication overhead by 1000 to 10000 times. As a result, each node can train independently without needing synchronization at every step, significantly lowering network latency and bandwidth demands. This makes it feasible for a wide range of hardware, from high-end RTX 4090 GPUs to enterprise-level A100 and H100 GPUs, to work together seamlessly. Solana plays a critical role in Psyche by serving as the coordination hub. Smart contracts on the Solana blockchain store training metadata, participant lists, and randomly assign tasks, ensuring that the entire training process is transparent, tamper-proof, and censorship-resistant. AIbase points out that Solana's high throughput and low transaction fees make it an ideal platform for managing the microtransactions involved in distributed AI training, enhancing Psyche's scalability across the globe. Open Source and Community-Driven Innovation Psyche is committed to open-source principles to promote AI democratization. All training data, checkpoints, and results will be freely available on platforms like Hugging Face and GitHub, allowing global developers to access them without cost. The code for Psyche is already open, and developers can participate in optimizations through a Rust-based system and P2P networks. They can also build custom applications using the provided documentation. Nous Research plans to further expand Psyche's capabilities, including support for model fine-tuning and community data input. This will enable smaller teams and even individuals to customize AI models for specific use cases, such as education, healthcare, or creative fields. For instance, community members can upload domain-specific data to train models tailored to their needs. Industry Impact: The Rise of Decentralized AI The launch of Psyche marks a significant milestone in the rapid development of decentralized AI. AIbase's analysis suggests that compared to traditional cloud-based AI, Psyche's distributed architecture not only reduces training costs but also transforms the contribution of computational power from a form of "donation" into a "transaction" through blockchain incentives, attracting widespread participation. In December 2024, Psyche successfully completed a test training of a 15-billion-parameter model over 11,000 steps, demonstrating its stability across a global network. In April 2025, Nous Research secured $50 million in Series A funding led by Paradigm, valuing the company at $1 billion. This investment has accelerated the development and expansion of Psyche's capabilities. Notably, the testnet launched just 44 minutes after receiving donations worth $500,000 in GPU compute power, underscoring the community's strong enthusiasm for Psyche. AIbase predicts that this community-driven model will spur global innovation, challenging the dominance of centralized AI giants like OpenAI and DeepSeek. A Turning Point for AI Democratization As a leading authority in AI, AIbase believes that the introduction of Psyche represents a turning point for both Nous Research and the decentralized AI movement. By combining DisTrO and Solana, this platform offers unprecedented opportunities for global developers to engage in AI development, breaking down technological and resource barriers. Of particular interest is the potential impact on China's AI community. Models like Qwen3 could benefit from Psyche's open-source framework by achieving more efficient local fine-tuning. This breakthrough not only empowers a diverse range of developers but also fosters a more equitable and inclusive ecosystem in the realm of artificial intelligence.