The Holy Grail of Crypto AI: Frontier Exploration of Decentralized Training

In the entire value chain of AI, model training is the most resource-intensive and technically challenging stage, directly determining the upper limit of the model's capabilities and its actual application effectiveness. Compared to the lightweight calls during the inference phase, the training process requires continuous large-scale computing power investment, complex data processing workflows, and high-intensity optimization algorithm support, making it the true "heavy industry" of AI system construction. From an architectural paradigm perspective, training methods can be divided into four categories: centralized training, distributed training, federated learning, and decentralized training, which is the focus of this article.

Centralized training is the most common traditional method, completed by a single institution within a local high-performance cluster, with all training processes coordinated by a unified control system, from hardware, underlying software, cluster scheduling systems, to all components of the training framework. This deep collaborative architecture maximizes the efficiency of memory sharing, gradient synchronization, and fault tolerance mechanisms, making it very suitable for training large-scale models such as GPT and Gemini, with advantages of high efficiency and controllable resources. However, it also faces issues such as data monopolies, resource barriers, energy consumption, and single-point risks.

Distributed training is the mainstream method for training large models. Its core is to decompose the model training tasks and distribute them to multiple machines for collaborative execution, in order to break through the computational and storage bottlenecks of a single machine. Although it has "distributed" characteristics in a physical sense, it is still controlled, scheduled, and synchronized by a centralized organization, often running in high-speed local area network environments. Through NVLink high-speed interconnect bus technology, the main node coordinates various sub-tasks uniformly. Mainstream methods include:

• Data parallelism: each node trains different data with shared parameters, requiring model weights to match.
• Model parallelism: Deploying different parts of the model on different nodes to achieve strong scalability.
• Pipeline parallelism: Phased serial execution to improve throughput
• Tensor Parallelism: Refined segmentation of matrix calculations to enhance parallel granularity

Distributed training is a combination of "centralized control + distributed execution", analogous to the same boss remotely directing multiple "office" employees to collaborate on completing tasks. Currently, almost all mainstream large models are trained using this method.

Decentralization training represents a more open and censorship-resistant future path. Its core feature lies in: multiple untrusted nodes collaborating to complete training tasks without a central coordinator, usually driven by protocols for task distribution and cooperation, and leveraging encryption incentive mechanisms to ensure the honesty of contributions. The main challenges faced by this model include:

• Heterogeneous devices and segmentation difficulties: Coordinating heterogeneous devices is challenging, and task segmentation efficiency is low.
• Communication efficiency bottleneck: network communication is unstable, and gradient synchronization bottleneck is significant.
• Lack of Trusted Execution: The lack of a trusted execution environment makes it difficult to verify whether nodes are truly participating in the computation.
• Lack of unified coordination: no central scheduler, complex task distribution and exception rollback mechanisms.

Decentralization training can be understood as: a group of global volunteers contributing computing power to collaboratively train models. However, "truly feasible large-scale Decentralization training" remains a systemic engineering challenge, involving multiple aspects such as system architecture, communication protocols, cryptographic security, economic mechanisms, and model verification. Whether it can achieve "collaborative effectiveness + incentive honesty + correct results" is still in the early prototype exploration stage.

Federated learning, as a transitional form between distributed and Decentralization, emphasizes local data retention and centralized aggregation of model parameters, making it suitable for scenarios that prioritize privacy compliance. Federated learning has the engineering structure of distributed training and local collaboration capabilities, while also possessing the advantage of data dispersion from Decentralization training. However, it still relies on trusted coordinators and does not have the characteristics of being fully open and resistant to censorship. It can be seen as a "controlled Decentralization" solution in privacy-compliant scenarios, relatively moderate in terms of training tasks, trust structure, and communication mechanisms, making it more suitable as a transitional deployment architecture in the industry.

Decentralization training boundaries, opportunities, and realistic paths

From the perspective of training paradigms, Decentralization training is not suitable for all types of tasks. In certain scenarios, due to the complexity of task structures, extremely high resource demands, or difficulties in collaboration, it is inherently unsuitable for efficient completion across heterogeneous, trustless nodes. For instance, large model training often relies on high memory, low latency, and high bandwidth, making it difficult to effectively partition and synchronize in an open network; tasks with strong data privacy and sovereignty restrictions are limited by legal compliance and ethical constraints, making open sharing impossible; while tasks lacking a foundation for collaborative incentives miss the external motivation for participation. These boundaries together constitute the current realistic limitations of Decentralization training.

However, this does not mean that Decentralization training is a pseudo-proposition. In fact, in task types that are lightweight, easy to parallelize, and incentivizable, Decentralization training shows clear application prospects. These include but are not limited to: LoRA fine-tuning, behavior alignment post-training tasks, data crowdsourcing training and labeling tasks, resource-controllable small foundation model training, and collaborative training scenarios involving edge devices. These tasks generally exhibit high parallelism, low coupling, and tolerance for heterogeneous computing power, making them very suitable for collaborative training through P2P networks, Swarm protocols, distributed optimizers, and other methods.

Decentralization training classic project analysis

Currently, the representative blockchain projects in the forefront of decentralized training and federated learning mainly include Prime Intellect, Pluralis.ai, Gensyn, Nous Research, and Flock.io. In terms of technological innovation and engineering implementation difficulty, Prime Intellect, Nous Research, and Pluralis.ai have proposed a number of original explorations in system architecture and algorithm design, representing the cutting-edge direction of current theoretical research; while Gensyn and Flock.io have relatively clear implementation paths, and preliminary engineering progress can already be seen. This article will sequentially analyze the core technologies and engineering architectures behind these five projects, and further explore their differences and complementary relationships in the decentralized AI training system.

Prime Intellect: A pioneer of verifiable training trajectories in reinforcement learning collaborative networks.

Prime Intellect is committed to building a trustless AI training network, allowing anyone to participate in training and receive credible rewards for their computational contributions. Prime Intellect aims to create a verifiable, open, and fully incentivized AI Decentralization training system through three major modules: PRIME-RL + TOPLOC + SHARDCAST.

01, Prime Intellect protocol stack structure and key module value

02, Detailed Explanation of Prime Intellect Training Key Mechanisms

#PRIME-RL: Decoupled Asynchronous Reinforcement Learning Task Architecture

PRIME-RL is a task modeling and execution framework customized by Prime Intellect for decentralized training scenarios, specifically designed for heterogeneous networks and asynchronous participation. It adopts reinforcement learning as the primary adaptation object, structurally decoupling the training, inference, and weight uploading processes, allowing each training node to independently complete task loops locally and collaborate through standardized interfaces with verification and aggregation mechanisms. Compared to traditional supervised learning processes, PRIME-RL is more suitable for achieving flexible training in environments without centralized scheduling, reducing system complexity and laying the groundwork for supporting multi-task parallelism and policy evolution.

#TOPLOC: Lightweight Training Behavior Verification Mechanism

TOPLOC is a core mechanism for training verifiability proposed by Prime Intellect, used to determine whether a node has truly completed effective policy learning based on observational data. Unlike heavyweight solutions like ZKML, TOPLOC does not rely on full model recomputation, but instead completes lightweight structural verification by analyzing the local consistency trajectories between "observation sequence ↔ policy update." It transforms the behavioral trajectories during the training process into verifiable objects for the first time, which is a key innovation for achieving trustless training reward distribution, providing a feasible path for building an auditable and incentivized Decentralization collaborative training network.

#SHARDCAST: Asynchronous Weight Aggregation and Propagation Protocol

SHARDCAST is a weight propagation and aggregation protocol designed by Prime Intellect, optimized for real network environments that are asynchronous, bandwidth-constrained, and with variable node states. It combines gossip propagation mechanisms with local synchronization strategies, allowing multiple nodes to continuously submit partial updates in an unsynchronized state, achieving progressive convergence of weights and multi-version evolution. Compared to centralized or synchronized AllReduce methods, SHARDCAST significantly enhances the scalability and fault tolerance of Decentralization training, serving as a core foundation for building stable weight consensus and continuous training iterations.

#OpenDiLoCo: Sparse Asynchronous Communication Framework

OpenDiLoCo is a communication optimization framework independently implemented and open-sourced by the Prime Intellect team based on the DiLoCo concept proposed by DeepMind. It is specifically designed to address common challenges in decentralized training such as bandwidth limitations, device heterogeneity, and unstable nodes. Its architecture is based on data parallelism, constructing sparse topological structures like Ring, Expander, and Small-World to avoid the high communication overhead of global synchronization, relying only on local neighboring nodes to achieve collaborative model training. By combining asynchronous updates with fault tolerance mechanisms, OpenDiLoCo enables consumer-grade GPUs and edge devices to stably participate in training tasks, significantly enhancing the accessibility of global collaborative training, making it one of the key communication infrastructures for building decentralized training networks.

#PCCL: Collaborative Communication Library

PCCL is a lightweight communication library tailored by Prime Intellect for decentralized AI training environments, aiming to address the adaptation bottlenecks of traditional communication libraries in heterogeneous devices and low-bandwidth networks. PCCL supports sparse topology, gradient compression, low-precision synchronization, and checkpoint recovery, and can run on consumer-grade GPUs and unstable nodes. It serves as a foundational component supporting the asynchronous communication capabilities of the OpenDiLoCo protocol. It significantly enhances the bandwidth tolerance and device compatibility of training networks, paving the way for building a truly open and trustless collaborative training network by bridging the "last mile" of communication infrastructure.

03, Prime Intellect Incentive Network and Role Division

Prime Intellect has built a permissionless, verifiable training network with economic incentives, allowing anyone to participate in tasks and receive rewards based on real contributions. The protocol operates based on three core roles:

• Task initiator: define training environment, initial model, reward function, and validation criteria
• Training Node: Execute local training, submit weight updates and observation trajectories
• Validation Node: Use the TOPLOC mechanism to verify the authenticity of training behavior and participate in reward calculation and strategy aggregation.

The core process of the protocol includes task publishing, node training, trajectory verification, weight aggregation, and reward distribution, forming an incentive closed loop centered around "real training behavior".

04, INTELLECT-2: The release of the first verifiable Decentralization training model.

Prime Intellect released INTELLECT-2 in May 2025, which is the world's first large-scale reinforcement learning model trained by asynchronous, trustless decentralized nodes, with a parameter scale of 32B. The INTELLECT-2 model was collaboratively trained by over 100 GPU heterogeneous nodes across three continents, using a fully asynchronous architecture, with a training duration exceeding 400 hours, demonstrating the feasibility and stability of asynchronous collaborative networks. This model not only represents a breakthrough in performance but also marks the first systematic implementation of Prime Intellect's proposed "training is consensus" paradigm. INTELLECT-2 integrates core protocol modules such as PRIME-RL, TOPLOC, and SHARDCAST, marking the first achievement of an open, verifiable, and economically incentivized closed-loop in decentralized training networks.

In terms of performance, INTELLECT-2 is based on QwQ-32B training and has undergone specialized RL training in code and mathematics, placing it at the forefront of current open-source RL fine-tuning models.