Fortytwo envisions a decentralized AI network—Swarm Inference—where hundreds of small models run on consumer devices to deliver scalable, potentially cheaper AI inference than centralized services.

A crypto-based incentive and reputation system rewards high-quality contributions, where top-performing nodes earn per-round rewards and gain reputation, while underperforming nodes lose standing.

Financial projections suggest node operators could earn roughly 10% more than comparable GPU rental services, with higher payouts for specialized tasks, such as CT scan analysis models.

Tests under extraneous information conditions show Swarm Inference maintains higher accuracy, avoiding the repetitive reasoning and detours seen in frontier models.

The design enables background operation with dynamic load-balancing so user tasks aren’t disrupted, fostering a global community of AI developers and researchers.

The network uses the Monad blockchain and FOR tokens to enable secure transactions, compensation, and decentralized governance.

The network coordinates many tiny, specialized models across a global device network, with operators selecting models and contributing to a collaborative, open AI ecosystem.

This approach aims to reduce costs, increase resilience and privacy, and minimize single points of failure by distributing computation and employing blockchain governance.

Swarm Inference uses hundreds of interconnected nodes hosting individual language models, coordinating to yield a single, high-quality answer.

Fortytwo plans open participation from node operators, model providers, and data scientists, and intends to release an API later this year to tackle demanding use cases like coding, deep research, and advanced reasoning.

Nodes can include open-source models like Qwen3-Coder and Gemma3, with Fortytwo’s Strand-Rust-Coder-14B, all operating as black boxes that only share inference results.

A primary advantage cited is cost efficiency, with swarm inference reportedly up to three times cheaper per token than frontier models, depending on task complexity.