Local AI Hardware.
Built for AI.

Local AI means the model runs on hardware you control instead of sending every request to a cloud service. For ClawBox, that means OpenClaw, local models, voice processing, and automations can run on the Jetson Orin Nano inside your home or office, with optional cloud providers available only when a task needs them.

Local AI hardware is a dedicated device that runs AI models (LLMs, speech recognition, image processing) entirely on your own hardware — no cloud, no internet required for AI inference. ClawBox is purpose-built local AI hardware with a 67 TOPS NVIDIA Jetson Orin Nano that runs models like Qwen2.5 3B at about 23 tokens/second and Phi-3 Mini at about 22 tokens/second, measured on a production unit.

Measured on a production ClawBox with Ollama: Llama 3.2 1B (~45 tok/s), Gemma 2 2B (~26 tok/s), Qwen2.5 3B (~23 tok/s), Phi-3 Mini 3.8B (~22 tok/s), plus Whisper STT (90+ languages, local) and Kokoro TTS. The 1-4B class is the sweet spot for the 8GB unified memory; 7-8B models load but run near 10 tok/s and leave little headroom alongside the assistant. Managed via Ollama, the de-facto local AI model manager.

A GPU desktop (e.g., RTX 4090) runs AI much faster, but costs €1,500-3,000+, draws 300-500W, and isn't designed for 24/7 always-on use as an AI assistant. ClawBox draws only 15W, costs €549, and is purpose-built for AI assistant workloads — with messaging integrations, browser automation, voice processing, and proactive monitoring built in. It's not about raw speed; it's about having an AI that's always on, always available, always private.

With ClawBox, no. Setup is 5 minutes: plug in the power, connect to WiFi, scan the QR code to link Telegram or WhatsApp. OpenClaw (pre-installed) handles all AI model management, voice processing, and integrations automatically. You don't need to install drivers, compile models, or configure software. For advanced customization, full SSH/root access is available.

To run local AI effectively, you need at least 8GB of unified or GPU memory, a modern CPU or GPU/NPU with AI acceleration, and an always-on device. ClawBox exceeds all these: 8GB LPDDR5 unified memory, 67 TOPS NVIDIA GPU, 1024 CUDA cores, and 512GB NVMe for model storage. See our full hardware requirements at clawbox.com/openclaw-hardware-requirements.