Gradient Network

Gradient Network is a decentralized AI infrastructure protocol built on Solana, providing an open layer for distributed model inference, reinforcement learning training, and peer-to-peer data transport. The project launched in 2024 and has evolved from a 1.6-million-node connectivity experiment into a full-stack AI compute platform, open-sourcing its core inference engine in late 2025. In June 2025, the Singapore-based team raised $10 million in seed funding led by Pantera Capital and Multicoin Capital, with participation from HSG (formerly Sequoia Capital China).

Background and Mission

Gradient was founded on the premise that AI infrastructure should be permissionless and broadly accessible rather than controlled by a handful of hyperscaler data centers. Co-founder Eric Yang has framed the project's goal as making intelligence a public good rather than a corporate asset. The network routes AI workloads across ordinary consumer devices — laptops, desktops, phones, and IoT hardware — coordinated through Solana's blockchain for settlement and payments.

The $10 million seed round funds development of two core protocols, Lattica and Parallax, and positions Gradient against other decentralized AI infrastructure efforts including SingularityNET, Bittensor, and Gensyn.

The Open Intelligence Stack

Gradient organizes its technology into what it calls the Open Intelligence Stack (OIS): three interconnected protocols covering data transport, model inference, and reinforcement learning training.

Lattica is the foundation layer — a universal peer-to-peer data communication protocol that moves model weights, inference tokens, and file shards across heterogeneous devices without centralized intermediaries. Lattica was built on connectivity data gathered during Gradient's early Sentry Node program, which mapped real-world network conditions across more than 190 countries. It handles the low-level networking that makes distributed inference and training practical at scale, supporting movement of data between nodes regardless of hardware type or geographic location.

Parallax is Gradient's distributed model inference engine. Its architecture has three layers: a Runtime Layer that orchestrates high-throughput LLM serving, a Communication Layer that manages peer tensor streaming across devices, and a Worker Layer that executes tasks on heterogeneous hardware. GPU workers run a modified version of SGLang; Apple Silicon workers use an MLX-optimized runtime. Parallax currently supports over 40 open-source models ranging from 0.6 billion parameters to trillion-parameter mixture-of-experts architectures, including DeepSeek-V3.2, DeepSeek-R1, Qwen3, MiniMax-M3, and Kimi-K2 variants.

Gradient open-sourced Parallax in October and November 2025, making it available for self-hosting and distributed participation. The team reports benchmarks showing 5.3× lower inter-token latency and 3.1× higher throughput compared to the Petals distributed inference framework. Integration partners include vLLM, SGLang, Qwen, and DeepSeek.

Echo is Gradient's distributed reinforcement learning framework. It uses a Dual Swarm Architecture that decouples inference — offloaded to the distributed Parallax network — from training, which runs on high-performance clusters. This separation lets large-scale reinforcement learning proceed without requiring every participating device to support full GPU training runs. Echo-2 launched in February 2026 with a claimed 80% reduction in AI post-training costs versus traditional cloud approaches. Published results from the Echo research program include a 7B math model that outperforms Qwen2.5-32B and a 30B reasoning model that outperforms DeepSeek-R1 on the Sokoban benchmark.

Gradient Cloud is the enterprise-facing access layer built on top of the full stack. It exposes OpenAI-compatible API endpoints and interactive model playgrounds, allowing businesses and developers to use distributed AI compute without operating their own node infrastructure.

Sentry Nodes: The Bootstrap Phase

Before Lattica and Parallax were operational, Gradient ran the Sentry Node Open Beta — a lightweight Chrome browser extension that participants installed to contribute connectivity data to the nascent network. Each Sentry Node consumed less than 1 MB of data per month and operated intermittently, measuring network conditions including NAT type, bandwidth, and latency across regions.

By the time the program concluded, over 1.6 million Sentry Nodes had been deployed across 190-plus countries, facilitating more than 1.6 billion peer-to-peer connections. Participants accumulated points through uptime rewards (0.15 points per 10 minutes of Season 1 uptime), tap rewards for successful peer-to-peer connections, and a 10% perpetual referral bonus on primary referees' earnings. Season 1 of the Sentry Node Open Beta concluded on August 28, 2025 — what Gradient describes as the program's graduation from connectivity testing into active compute contribution.

The connectivity map assembled during the Sentry Node phase became the empirical foundation for Lattica's design.

Worker Nodes and Participation

With Sentry Nodes retired, participation centers on the Edge Host Pilot Program for worker nodes running Parallax. Node performance is scored using the formula S = N × H × C, where N (Network Score) evaluates NAT type, bandwidth, and latency on a 1–5 scale; H (Hardware Score) rates CPU cores, GPU tier, RAM capacity, and storage type; and C (Consistency Score) tracks a 7-day moving average of uptime and liveliness, with thresholds at 75–90% and above 90% uptime.

Worker nodes earn two types of rewards: Proof of Connectivity (POC) rewards for network availability, and Proof of Work (POW) rewards for completed inference tasks. Account tier boosts — spanning Iron through Diamond tiers — add 0–10% multipliers on top of base earnings.

Solana Integration

Gradient uses Solana as the coordination and settlement layer for its distributed compute network. Node rewards and inter-device payments settle on-chain, with Solana's high throughput and low transaction costs enabling near real-time settlement. Gradient has noted that this settlement architecture gives customers a meaningful advantage over traditional cloud platforms for workloads where payment coordination is tightly coupled to compute delivery.

Token

As of mid-2026, Gradient has not announced a token generation event or published tokenomics. The points system from the Sentry Node program and ongoing worker node participation are widely expected to underlie a future airdrop, but no official date, allocation, or TGE timeline has been confirmed. The project originally signaled a potential Q4 2025 launch through community channels; that timeline was not met.