Acrab, a developer of agentic AI computing infrastructure, has introduced its first-generation GΞLIX 1 edge AI system-on-chip together with Agent Box, a personal edge AI system built on the company’s computing platform. The company said the SoC is manufactured using a 5 nm process and designed to support local inference for open-source AI models in the 100-billion-parameter class.
The GΞLIX 1 architecture combines CPU, GPU and NPU resources with a unified memory design developed for large AI models and agentic workloads. The SoC incorporates a 20-core Arm CPU, multicore NPU acceleration and 273 GB/s of unified memory bandwidth. The company said coordinated execution across the CPU, GPU and NPU is intended to address the computing performance, memory bandwidth and power requirements associated with running large models locally rather than relying entirely on cloud infrastructure.
Acrab designed the SoC to reduce the delay before a model begins responding, particularly when processing long prompts and large context windows. In company testing, GΞLIX 1 achieved a prefill rate of 1,416.8 tokens per second using a Gemma 26B A4B configuration with a 40K KV cache and a 10K-token input. The company compared that result with a prefill rate of 188.9 tokens per second on an Apple Mac mini with an M4 Pro processor and reported as much as 7.5 times faster prefill performance under the stated test conditions. The results apply to the specific model, system and workload configuration used in the company’s testing.
The GΞLIX 1 is Acrab’s first-generation edge AI system-on-chip. The company said the device is manufactured using a 5 nm process and designed to support local large-model inference.
“Generative AI helped people find answers. Agentic AI will help them get things done,” said Dr. Ken Phua, CEO of Acrab. “Running models in the 100 billion parameter class on a system small enough to sit on a desk presents a significant computing challenge. GΞLIX 1 is designed to deliver the performance, memory bandwidth and responsive local inference required, while Agent Box shows how that capability can become a complete user experience.”
Agent Box demonstrates how the GΞLIX 1 SoC can be integrated with Acrab’s software, system architecture and agentic AI capabilities. The personal edge AI system is designed to support local language and vision model inference, multimodal interaction, persistent memory and the orchestration of agents, systems and connected devices. The company said user data and memories can remain stored locally, while cloud resources can still be used when required by a particular task. Processing AI workloads locally can reduce the delays associated with sending data to a remote service and may also reduce dependence on metered cloud inference. The actual performance, privacy and cost benefits will depend on the model, system configuration, application and balance between local and cloud processing.
In addition to the SoC, Acrab has developed an optimized runtime, developer toolchain, reference designs and agent operating system capabilities. These software and system layers are intended to help developers create products that can retain context, coordinate tasks and interact with connected equipment. Agent Box is the first product implementation of the broader platform and demonstrates the integration of Acrab’s silicon and software within a complete edge AI system. The company plans to work with device manufacturers and developers to integrate its computing platform into AI network-attached storage systems, AI PCs, smart vehicles, industrial robots and service robots. It is positioning the platform as a common computing foundation for agentic AI applications across personal, automotive and industrial environments.
“Our goal is to give device makers and developers the foundation to bring agentic intelligence into many different products and environments,” Dr. Phua said. “Agent Box demonstrates what the technology can do today, while GΞLIX 1 and our full-stack platform are designed to support a much broader ecosystem of devices and applications.”
Founded in 2024 and headquartered in Singapore, Acrab develops computing architecture and integrated software platforms for agentic AI systems. In June 2026, the company announced that it had received more than $350 million in cumulative financing and identified its investors as including Vertex Ventures Southeast Asia & India, Vertex Growth and K3.
About Acrab
Acrab develops computing infrastructure for agentic AI systems. Its technology combines purpose-designed silicon, edge AI models, full-stack software and system orchestration to support personalized assistance and real-time execution across a range of edge environments. For more information, please click here.
(Editor’s Note: All trademarks mentioned in this article, including company names, product names, and logos, are the property of their respective owners. Use of these trademarks is for informational purposes only and does not imply any endorsement.)
Molly Bakewell Chamberlin is the Founder, Publisher and Editor-in-Chief of Components Source Network, a 100% woman- and disability-owned B2B publishing and expert marketplace platform serving the global high-tech manufacturing and engineering community. She is the architect, author and developer of the Network’s 11 specialized e-news sites and 11 peer-vetted expert e-marketplaces, including Semiconductor Industry Monthly. With more than 25 years of experience in strategic technical marketing, public relations and business development, she is also the Founder and President of Embassy Global, LLC, which has supported more than 200 high-tech industry brands worldwide since 2008. Molly also serves pro bono as Second Vice Chair of the Government Team for the WBENC Forum and is a long-standing advocate for small business, supplier diversity, sustainability and women in STEM. Connect with Molly on LinkedIn.
- Agentrys Raises $24.5M for AI Chip Design Automation - August 28, 2026
- ACM Research Expands Wet Processing Platform - August 26, 2026
- Lam Research Plans $3B Global R&D Lab Expansion - August 26, 2026