Singapore startup Acrab has launched Agent Box, a mini PC designed to execute artificial intelligence models with up to 100 billion parameters locally. The desktop device operates without an internet connection, offering developers and users an alternative to cloud-dependent generative AI platforms.
Running AI models directly on a computer eliminates token consumption costs while protecting user privacy and enabling personalized digital assistants with persistent memory. Major chipmakers such as AMD and NVIDIA already offer hardware for local AI processing, though their systems carry high price tags.
At the center of Agent Box is a custom-developed processor called G≡LIX 1, manufactured using a 5-nanometer process built for demanding workloads. According to Acrab, the machine delivers performance close to expensive professional workstations such as the NVIDIA DGX Spark, which costs 4,800 euros. The startup claims Agent Box achieves this level of performance at roughly one-fifth of the cost while consuming half the energy.
Internal benchmark tests conducted by the company showed that the mini PC outperformed an Apple Mac mini equipped with an M4 Pro chip. When processing large models, Agent Box exceeded the Mac mini's input processing speed by more than seven times.
Hardware architecture and memory specifications
The G≡LIX 1 chip combines a 20-core ARM CPU, a graphics processing unit rated at 3 TFLOPS, and a multi-core neural processing unit specifically optimized for language model inference. The system reaches a total computing performance of 700 TOPS, supported by unified memory with a bandwidth of 273 GB/s. For comparison, Apple's M5 chip provides a unified memory bandwidth of 153 GB/s.
Alongside the hardware, Acrab developed a software platform that includes AI runtime environments, developer tools, and capabilities for autonomous agents. Company demonstrations showed voice commands initiating the automatic creation and 3D printing of models, operating robotic vacuums, and controlling smart home devices such as lights and electronic locks without sending data to cloud servers.
Expansion plans and industrial applications
Acrab aims to establish a framework for third-party manufacturers to integrate autonomous intelligence into their own hardware products. The Singapore startup plans to bring its technology to data storage systems, industrial robotics, and electric vehicles.
