Fortytwo, a startup based in Silicon Valley, was established last year with the innovative concept that a decentralized network of smaller AI models operating on personal computers can offer both scalability and cost benefits compared to traditional centralized AI services. Recently, the company released benchmark findings that suggest its swarm inference method surpasses the performance of leading AI models, including OpenAI“s GPT-5, Google“s Gemini 2.5 Pro, Anthropic“s Claude Opus 4.1, and DeepSeek R1 in various reasoning assessments.
The company”s reasoning tests included metrics like GPQA Diamond, MATH-500, AIME 2024, and LiveCodeBench. According to Fortytwo, the swarm inference approach mitigates accuracy issues often encountered by large models when engaging in complex reasoning tasks. These larger models may fall into reasoning loops, whereas swarm inference leverages multiple smaller models, ranks their outputs, and generates a refined response.
Cost-effectiveness is another significant advantage, as Fortytwo”s system operates on distributed consumer hardware rather than relying on expensive data centers. “Inference through the swarm is up to three times cheaper than frontier reasoning models from OpenAI and Anthropic on a per-token basis,” stated Ivan Nikitin, co-founder and CEO, in correspondence with The Register. “Actual cost varies based on the complexity of the task.”
In a recent interview, Nikitin explained that the motivation behind pursuing decentralization was not merely to innovate, but to address the increasing scarcity of centralized computing resources. He highlighted how past AI projects faced usage limits, an issue exacerbated by the rising demand for coding models. Developers utilizing coding AI have struggled to make sufficient requests to meet their project requirements.
“The centralized AI industry is racing towards multi-billion-dollar contracts for new data centers and other infrastructure,” Nikitin remarked. “We believe this approach is not sustainable, as demand will always outpace supply.” He noted that the rising complexity of multistep reasoning tasks will necessitate even greater computational power.
Nikitin and his co-founders recognized that many individuals have untapped computing power in their home desktops, which are often more capable than needed for everyday tasks. Additionally, advancements in AI have demonstrated that smaller, specialized models can outperform expensive large models in specific domains. “We aimed to combine these factors to establish a network that deploys specialized models to work synergistically,” Nikitin explained.
In a preprint paper titled “Self-Supervised Inference of Agents in Trustless Environments,” released via ArXiv, Fortytwo”s decentralized network connects numerous small language models (SLMs), including open-source models and their proprietary models like Strand-Rust-Coder-14B. Each network node operates independently, allowing node operators to run any privately developed or downloaded model without exposing it to the network. Only the inferences are shared, ensuring data privacy.
However, latency remains a challenge. “Fortytwo prioritizes quality over speed,” Nikitin noted. The swarm inference process may introduce a delay of approximately 10 to 15 seconds, as multiple nodes collaborate and evaluate their outputs before reaching a consensus.
Privacy concerns are also present, although they may be less pronounced than with large AI firms that aggregate data for advertising purposes. On Fortytwo”s decentralized network, knowledgeable node operators could potentially access prompts and responses for locally running models, but this data volume would be significantly lower than what major companies like Anthropic, Google, or OpenAI collect.
Fortytwo is exploring strategies to enhance privacy, including adding noise to prompts, and has partnered with Acurast, a decentralized computing network designed for mobile devices. Nikitin indicated that mobile phones typically possess stronger Trusted Execution Environments than desktop systems, which could facilitate private inference.
The company”s vision is to foster an open community that allows machine learning engineers and data scientists to contribute to advanced AI technologies without the need for substantial financial backing. Participants will have the chance to develop specialized models in particular fields and receive compensation for their contributions. Once the commercial phase begins, node operators will run local AI models and receive potential compensation in cryptocurrency.
Customers using the API will pay the service provider in fiat currency, which will then be distributed to the Fortytwo network. The network will allocate a portion of the funds in Fortytwo Network FOR tokens to node operators based on the requests they fulfill. “Crypto is vital for creating a system without gatekeeping,” Nikitin explained, emphasizing the importance of maintaining decentralized reputation management.
In the current model, not all participants receive compensation. Inference rounds involve multiple contributions, and only the top-performing nodes receive rewards, while those that fail to provide relevant answers risk losing reputation points, incentivizing continuous improvement.
Daily engagement in Fortytwo”s network through the Devnet Program ranges from approximately 200 to 800 computers. The dashboard indicates that the company has distributed over 145 million FOR tokens, with the exact fiat value to be determined post-launch. The network is currently functioning on the Monad Testnet, where assets hold no redeemable value.
While it is too early to determine potential earnings for network participants, Nikitin believes the goal is to provide better returns than existing services like VastAI. “In our case, nodes can operate in the background without disrupting daily tasks,” he added. “Participants can expect earnings that cover electricity costs and offer some passive income.” For node operators running specialized models, income could reach significant amounts based on simulations from the previous year.
Nikitin concluded that Fortytwo aims to provide a reliable inference backend for various demanding tasks, including reasoning, coding, and medical analysis. The API is designed for seamless integration into mobile and web applications, similar to established AI services.
The architecture of Fortytwo”s network is designed to utilize idle computing power, ensuring minimal disruption to users” daily activities. “We have implemented a dynamic load-balancing system that optimizes node activity during lighter tasks,” Nikitin stated. “Our ambition is to spark a grassroots movement among enthusiasts and professionals alike, encouraging them to develop and integrate their own models into the network.”
