Gonka Network Architecture: Sprint, Transfer Agents, DiLoCo

Transfer Agents are a key component of Gonka's architecture, acting as an intelligent gateway between clients and ML-nodes. When a user sends an AI request, it doesn't go directly to a GPU, but to a Transfer Agent—a specialized intermediary node that decides which GPU will process the request.

The process is as follows: the client makes a standard POST /v1/chat/completions request via the OpenAI-compatible API. The Transfer Agent verifies the cryptographic signature of the request, determines the necessary model, and finds an available ML-node with suitable characteristics. Each ML-node, upon registration, publishes its parameters: what models it supports, VRAM capacity, current bandwidth, and load. The Transfer Agent uses this data for load balancing—tasks are distributed evenly, rather than accumulating on a single node.

For fault tolerance, several Transfer Agents operate simultaneously in the network. If one fails, the client automatically switches to another. Each Transfer Agent publishes its address via the endpoint /v1/identity, allowing nodes and clients to dynamically discover each other. Transfer Agents also manage request queues: if all nodes are busy, the request is queued with priority based on the commission. This architecture resembles a CDN, but for AI computations—distributed, fault-tolerant, and without a single point of control.

Sprint is Transformer PoW 2.0, Gonka’s unique consensus that fundamentally differs from all existing blockchain protocols. In Bitcoin, miners spend 26 GW of power computing meaningless SHA-256 hashes — whose sole purpose is to prove that energy was spent. In Ethereum Proof of Stake, computational work is abandoned altogether — validators simply lock tokens, sacrificing decentralization for energy efficiency. Sprint offers a third way.

In Sprint, every computation is a real AI request. A user sends a prompt “write a Python function” → the GPU generates a response via the Qwen3-235B neural network → this inference simultaneously serves the user and confirms a block on the blockchain. The result: 99% of network resources go to useful work (AI inference), and only 1% to cryptographic security. For comparison: in Bitcoin, 100% of energy goes to security, 0% to useful work.

The network's operation is organized into epochs. In each epoch, Transfer Agents distribute AI tasks among ML-nodes. At the end of an epoch, a block is formed containing proofs of completed work. Rewards are distributed proportionally to each node's contribution — the more requests a GPU processed, the more GNK it receives. This creates a market incentive: hosts compete for tasks, optimizing performance and reducing the cost of inference for users.

DiLoCo is a technology for distributed AI model training that solves a fundamental problem: how to train a neural network on billions of parameters when GPUs are in different countries and connected by ordinary internet, not high-speed NVLink within a single data center?

The traditional approach to training requires parameter synchronization after each step—this is only possible with communication speeds of hundreds of gigabits/s, i.e., within a single NVIDIA cluster. DiLoCo rethinks the process: nodes synchronize parameters once every ~1000 steps, not after each. Between synchronizations, each node trains locally on its subset of data. This reduces bandwidth requirements by three orders of magnitude, making training over the internet practically feasible.

Optimization works on two levels: locally, each node uses AdamW—a standard optimizer for transformers. Globally, during synchronization, Nesterov momentum is applied—an algorithm that 'predicts' the direction of update and accelerates convergence. The result: models with 30–50 billion parameters can be trained on clusters of 8xH100, distributed across the globe, without a central server. For comparison: training GPT-4 required thousands of GPUs in a single data center with billions of dollars in investment. DiLoCo potentially allows achieving a comparable result on Gonka's distributed infrastructure.

Why is this important? Training is the most expensive part of AI. Companies like OpenAI spend hundreds of millions on a single training cycle. DiLoCo allows Gonka to eventually train its own models using the network's resources—without the need to build data centers worth billions. This makes Gonka not just an inference network, but a full-fledged AI platform with vertical integration.

PoC V2 is a verification mechanism that guarantees each ML-node actually performed the computation and didn't return random junk. This is critically important: without verification, an attacker could register a "node" that gives fake answers and receives rewards without spending a single watt on the GPU.

The mechanism works through cross-verification. The network randomly selects 1–10% of tasks and sends them for re-execution by another node. If the results match, both nodes receive a reward. If the results differ, an arbitration process (dispute) begins. The losing node loses 20% of its stake, which is distributed among honest participants. This penalty makes fraud economically unviable: the potential income from fake answers is significantly less than the risk of losing the stake.

Verification speed is ensured by BLS signatures—a cryptographic primitive that allows aggregating multiple signatures into one and verifying it in less than 10 milliseconds. This means that honesty checking does not slow down the network—the user receives a response without delay, and verification occurs in parallel.

For model training tasks (via DiLoCo), an additional mechanism is used—Proof-of-Learning. Each node records model weight hashes and optimizer states at each checkpoint on the blockchain. This creates an immutable audit trail: anyone can verify that training actually occurred and weights were not tampered with. Such two-level verification—PoC V2 for inference, Proof-of-Learning for training—makes Gonka one of the most secure decentralized AI networks, audited by CertiK.