Fortytwo is a infrastructure provider for AI agents that require higher reliability than a single model can provide. Their "Prime" product is designed specifically for agentic workflows where accuracy is a hard requirement, using a parallel-processing "bracket" system to verify outputs from multiple models.
In the broader ecosystem, Fortytwo acts as a bridge between decentralized compute and practical agent development. Their support for the Model Context Protocol (MCP) allows agents to connect to their decentralized swarm as easily as they would to a centralized API. By moving inference away from corporate-controlled data centers and into a peer-ranked network, they offer a path for developers to build agents that are less dependent on a single provider's uptime or policy changes.
Fortytwo is an AI research lab focused on an architectural shift from monolithic models to decentralized collectives. Founded in 2024 by Ivan Nikitin, Vladyslav Larin, and Alex Firsov, the company is developing what it calls "Swarm Inference." This method coordinates a network of small language models (SLMs) to achieve reasoning capabilities that match or exceed frontier models like GPT-4 or Claude. The project raised a $2.3 million pre-seed round in March 2025 from investors including Big Brain Holdings and CMT Digital to scale this peer-to-peer infrastructure.
The technical premise is that centralized AI faces a reliability bottleneck. Fortytwo addresses this by having multiple nodes in its decentralized network cross-check answers. When a query enters the network, several nodes generate independent responses, which are then compared and evaluated by the swarm. This peer-ranked consensus is designed to reduce hallucinations and ensure the final output is the most accurate result the collective can produce. Currently, the network reports over 3,500 active nodes and agents, with thousands of inferences completed daily.
Part of Fortytwo's strategy involves the creation of domain-specific expert models. Rather than training a general-purpose giant, the lab uses the swarm to generate high-quality synthetic datasets that train smaller, specialized models. A primary example is "Strand Rust Coder 14B," which the company claims is the best model for Rust programming. This model was trained on community-created synthetic data and then integrated back into the network to provide specific expertise to the broader swarm.
This self-evolving cycle is intended to make the network more capable as it grows. Every new model added to the system acts as a new expert node, contributing to the collective intelligence. The company reports that its swarm approach has already achieved top rankings on benchmarks such as AIME 2024 and LiveCodeBench, occasionally outperforming proprietary models from OpenAI and Google in raw accuracy tests.
The infrastructure for this network is built on the Monad blockchain, chosen for its high throughput which is necessary for handling the rapid consensus rounds required for swarm inference. By leveraging idle compute on consumer devices—Windows, Apple, and Linux—Fortytwo avoids the massive capital expenditures associated with centralized data centers. Node operators are incentivized to contribute their resources, earning points and future tokens for their computational work.
For developers, the primary point of entry is the Fortytwo Prime service. This is an inference API specifically built for AI agents that require high reliability. By running queries across a bracket of top frontier models in parallel and selecting the strongest answer, Prime acts as a high-assurance layer. The company is also an early adopter of the Model Context Protocol (MCP), making it easy for developers to integrate this swarm-based intelligence into existing agentic workflows without replacing their entire stack.
A high-reliability inference service for AI agents using parallel frontier models.
Fortytwo is hiring