Acurast brings real time AI decision model Laya to smartphone network

Acurast has deployed open source AI decision model Laya across its decentralized smartphone compute network, allowing the model to process real time decisions on Android devices instead of centralized data centers.

Summary

• Acurast has deployed open source Laya AI across its decentralized smartphone compute network.
• Laya processes structured decisions for gaming, classification, moderation and security tasks instead of generating conversational text.
• Acurast said decisions typically take around 0.2 to 1 second and are processed using smartphone CPUs without dedicated GPUs.
• More than 280,000 smartphones across over 175 countries have joined the Acurast network, according to the company.

According to Acurast the deployment runs Laya entirely on mobile edge nodes and uses the network's distributed smartphones to handle workloads including classification, gaming, navigation, security and content moderation.

Acurast brings Laya AI model to smartphone compute network

Laya, developed by Convai Innovations under an Apache 2.0 license, is designed to make structured decisions instead of generating conversational responses.

Given a defined state and a set of possible actions, the model evaluates the input and selects an option. Its open source repository describes three main decision formats covering choices between defined options, scoring against a set range and boolean decisions.

The model uses a non autoregressive architecture, meaning it does not generate an answer token by token in the way large language models typically produce text. Convai Innovations describes Laya as a "System 1" decision model built for tasks where an application needs a quick structured response.

Acurast is using the model to demonstrate how such workloads can operate across commodity mobile hardware. Each decision in its deployment is processed on an Android smartphone connected to the network, according to the company.

"By running Laya on Acurast, we've changed that," Acurast founder Alessandro De Carli said. "We are proving that System 1 AI can run securely and at scale on hardware everyone already owns, providing a truly open, decentralized alternative to centralized cloud lock-in."

Laya's public code and model weights are available under the Apache 2.0 license. Its repository lists a 421 million parameter English checkpoint, alongside separate models for multilingual and typed decision workloads.

What can Laya do on Acurast?

Acurast has set up several live demonstrations to test the model across different types of decision based tasks.

For gaming, Laya has been used to play Snake and Tetris by choosing movements and determining where falling pieces should be placed. A Doom demonstration requires the model to make decisions around movement, aiming, shooting, collecting items and exploring the game environment.

Other tests are closer to potential software applications. The model can classify emails as inbox messages, spam or phishing attempts and sort headlines into categories including news, satire, clickbait and manipulation. Acurast said it has tested Laya for detecting prompt injection attempts against AI assistants and identifying toxic messages in live chats.

Users can submit their own letters, headlines, prompts and other inputs to the demonstrations and observe the decisions made by the model.

Laya's developers have documented limitations as well. The project's repository says its general model performs poorly on some zero shot typed decision tests and recommends specializing the model for specific tasks. It advises keeping choice questions below roughly 20 options and recalibrating probabilities using application specific data.

Smartphones handle the AI workloads without GPUs

Acurast said decisions in its Laya deployment typically take roughly 0.2 to 1 second and are processed using smartphone CPUs rather than dedicated GPUs.

Work is distributed dynamically between devices participating in the network. Results come with cryptographic proof, while smartphone operators receive ACU, Acurast's native token, for providing processing capacity.

The network uses Trusted Execution Environments built into modern smartphones to isolate workloads and protect data while computation takes place, according to Acurast.

Acurast's approach relies on existing consumer hardware instead of requiring operators to purchase dedicated servers or GPUs. The company has previously applied the same model to blockchain infrastructure and other decentralized computing workloads.

Crypto.news previously reported that Acurast raised $5.4 million in May 2025 to develop its smartphone powered decentralized cloud network. At the time, the network had more than 72,000 smartphones and had processed 256 million transactions.

Acurast now says more than 280,000 smartphones across over 175 countries have joined the network, while the number of on chain transactions processed has passed 918 million.

The model fits into a growing effort to use consumer devices for decentralized AI workloads. Gaia introduced an AI smartphone in September 2025 that was designed to run AI locally on the device while letting users contribute compute to its decentralized network.

Acurast targets smaller real time AI workloads

Acurast is positioning the Laya deployment around workloads that need repeated decisions instead of long generated responses.

The distinction allows smaller models to handle tasks such as routing, moderation and classification without sending every request to a large generative model hosted by a centralized provider. Developers can deploy Laya themselves because its model weights and software are publicly available.

Convai Innovations' documentation describes Laya as a family of models rather than a single checkpoint. Its English version is based on ModernBERT large with 421 million parameters, while a multilingual version uses a 322 million parameter model and supports more than 100 languages. A separate checkpoint has been trained for typed decision workflows including customer service, invoice processing and security incidents.

Acurast said its implementation lets developers run those types of continuously operating workloads without obtaining API access or server capacity from a centralized cloud provider.

"Our goal is to open developers' eyes to the quite incredible opportunities that come from using Acurast compute," De Carli said. "This deployment proves that decision-oriented AI can run cheaply, verifiably, and without a data center in sight on a decentralized smartphone network today."

The company has made the Laya demonstrations available for users to submit their own inputs and watch the resulting decisions being processed on its mobile edge network.