deputyOS on Raspberry Pi 5
Run a personal AI assistant on a Raspberry Pi 5. Flash, boot, and finish the wizard in about 15 minutes — no SSH, no terminal.
- Setup time
- ~15 min
- Difficulty
- Easy
- Cost
- ~$80 one-off
- Acceleration
- native arm64
The Pi 5 is the recommended dedicated target. With 8 GB it runs the standard tier comfortably; the 16 GB board handles the rich tier including local models.
What you need
- Raspberry Pi 5 (8 GB or 16 GB recommended; 4 GB works on the lean tier)
- A quality microSD card (32 GB+) or, better, an NVMe HAT and SSD
- USB-C power supply — the official 27 W one, since undervolting causes throttling that looks like software trouble
- Ethernet for setup (Wi-Fi works, but wired makes first boot predictable)
Flash it
Pick rpi5 in the picker, download the .img.xz, and write it with
Raspberry Pi Imager or:
xzcat deputyos-rpi5-standard.img.xz | sudo dd of=/dev/sdX bs=4M status=progress conv=fsyncVerify the checksum against the signed manifest before writing. The picker links both the SHA256 and the minisign signature.
First boot
Insert the card, connect Ethernet, power on. First boot expands the filesystem and starts the wizard; give it two or three minutes.
Then open http://deputyos.local from any machine on the same network. If mDNS
does not resolve — common on guest and enterprise networks — find the address in
your router’s DHCP table instead. A QR code also appears on HDMI if you have a
screen attached.
The wizard covers profile choice, model provider credentials, and optionally linking the device to your account for tunnelled remote access and backup.
Storage
SD cards are the usual cause of Pi grief. They wear out, and the failure presents as random corruption rather than an obvious fault. For anything you care about, boot from NVMe via a HAT — it is faster and dramatically more durable.
Tier guidance
| RAM | Tier | What runs |
|---|---|---|
| 4 GB | lean | Cloud model providers; on-demand ClamAV; no local LLM |
| 8 GB | standard | Comfortable general use, small local models |
| 16 GB | rich | Local models at useful sizes, voice, heavier recall |
Known constraints
Local inference on the Pi 5 is CPU-only — there is no usable GPU offload — so expect a few tokens per second on small quantized models. Most Pi deployments use a cloud provider for the model and keep the Pi as the always-on assistant.