A Singapore-based startup backed by the world's largest battery manufacturer has unveiled a compact desktop computer designed to run large language models locally at a fraction of the cost of Nvidia's flagship AI appliance. Acrab's Agent Box, powered by a proprietary chip called G≡LIX 1, aims to bring agentic AI workloads to the edge without relying on cloud services.

What You Need to Know

The Agent Box is a mini PC that can run 100 billion parameter language models entirely on local hardware. Its custom G≡LIX 1 processor delivers 700 TOPS of AI performance, matching the capability of Nvidia's $5,000 DGX Spark while costing roughly one-fifth as much and using half the power. The startup behind the device, Acrab, appears to have financial backing from CATL, the world's largest electric vehicle battery maker.

The Agent Box Hardware

The G≡LIX 1 system-on-chip combines a 20-core Arm CPU, a 3 TFLOPS GPU, and a multicore neural processing unit optimized for large language model inference. The chip accesses 273GB/s of high-bandwidth unified memory and includes an 8MB L1 cache alongside a 768GB/s L2 cache. Acrab says the processor reaches 700 TOPS while keeping the entire system small enough to sit on a desk.

In internal testing, the Agent Box achieved a prefill rate of 1,416.8 tokens per second under a Gemma 26B A4B configuration. That compares with 188.9 tokens per second on an Apple Mac Mini M4 Pro, representing a claimed 7.5 times improvement in prefill speed over conventional consumer hardware. The NPU benefits from VAE vector acceleration that reportedly speeds up attention processing by up to three times.

  • G≡LIX 1 chip: Manufactured on a 5nm process, integrates a video DSP cluster, camera input subsystems, and a display engine within a single architecture.
  • 700 TOPS performance: Delivers inference speeds that compete with Nvidia's DGX Spark at a fraction of the price and power draw.
  • Local AI inference: Runs 100 billion parameter models entirely on-device, eliminating the need for cloud connectivity.

Software Ecosystem and CATL Backing

Beyond the hardware, Acrab provides a complete software platform with AI runtimes, developer toolchains, an operating system, reference designs, and orchestration tools for autonomous agents. Demonstrations showed the Agent Box handling voice commands to create and print 3D models, operate robotic vacuum cleaners, and control smart home devices including lights, air conditioners, and locks.

The company plans to extend its computing platform into AI NAS systems, industrial robots, service robots, and smart vehicles through hardware partnerships. Copyright notices accompanying the announcement reference CATL, the world's largest electric vehicle battery manufacturer. That relationship, if it reflects financial backing, could provide Acrab with significant resources to scale production and compete with established players.

Why This Matters

The Agent Box represents a direct challenge to Nvidia's dominance in the AI hardware market, which has long relied on expensive, power-hungry GPUs for both training and inference. By offering comparable performance at one-fifth the cost and half the power consumption, Acrab's device could accelerate the shift toward edge AI where models run locally rather than in the cloud. For developers and enterprises, this means lower operating costs, improved data privacy, and the ability to deploy AI agents in environments with limited or intermittent internet connectivity. CATL's potential involvement signals that a major industrial player sees value in moving AI inference to the edge, which could spur further investment and competition in this segment.