The GΞLIX 1 is built on a 5-nanometer process and integrates a 20-core Arm CPU, multicore NPU, and high-speed unified memory architecture. This hardware configuration is specifically tuned for large model inference, delivering 273 GB/s of memory bandwidth. In company benchmarks, the chip achieved a prefill rate of 1416.8 tokens per second using a Gemma 26B configuration, a performance metric Acrab claims is 7.5 times faster than the Mac Mini M4 Pro.
Dr. Ken Phua, CEO of Acrab, emphasizes that the shift from generative AI to agentic AI requires local processing to ensure privacy and low-latency interaction. By hosting persistent memory and orchestration layers on the device, the Agent Box aims to function as an always-on assistant that grows more customized as it accumulates user context. The company intends to license this full-stack platform, including the silicon and software runtime, to manufacturers of smart vehicles, industrial robots, and AI-enabled PCs. This strategy positions Acrab as an infrastructure provider, attempting to standardize how edge devices handle complex, goal-oriented tasks without relying on recurring cloud token fees.
Comments (0)
No comments yet. Be the first!