Platform

Hardware

Specifications below are limited to the public journal. A reference design is not a count of machines in operation.

commissionedPublished record

First commissioned node

One Raspberry Pi 5 worker commissioned on 24 October 2026, as recorded in Part 2.

Source · Building KRXY, Parts 2 and 6. Commissioning and single-session evidence, not a general benchmark.

commissionedPublished record

Second node, CPU only

A second Raspberry Pi 5 that joined on 28 October 2026, as recorded in Part 6. It has no Hailo accelerator, so it advertises CPU deployments only.

Source · Building KRXY, Part 6. Single-session evidence, not a general benchmark.

reference designPublished record

Reference node design

The planned shape of a node from Part 1, written before the hardware had been tried.

Source · Building KRXY, Part 1. A plan, not an inventory of running machines.

Specification comparison

Side by side, as each source describes it. An empty cell means the source does not say.

Hardware specifications by node
Specification	First commissioned nodecommissioned	Second node, CPU onlycommissioned	Reference node designreference design
Configuration	Raspberry Pi 5 with 16 GB of RAM. The journal records that it booted from USB storage and kept the same identity across a reboot.	Raspberry Pi 5 with 16 GB of RAM, running llama.cpp. Part 6 lists the Qwen2.5 1.5B Q4_K_M GGUF as available when its CPU service was ready and the 3B Q4_K_M GGUF as brought up for the two-node tests.	Raspberry Pi 5, 16 GB RAM, for every intended node.
Accelerator	Raspberry Pi AI HAT+ 2, named in Part 6 as the Hailo-10H accelerator. During commissioning the Hailo device was visible and firmware 5.1.1 was reported. Manufacturer performance figures were not treated as measurements.	None. Without Hailo hardware it cannot run the compiled HEF deployment.	Raspberry Pi AI HAT+ 2 on one machine only. The journal quotes a manufacturer rating of 40 TOPS INT4 and 8 GB of onboard memory, and says both were still unmeasured.
Storage	USB NVMe. The HAT uses the Pi 5 PCIe link, so the SSD could not use that same connection. Part 1's reference design describes 256 GB.	A roughly 256 GB USB-attached SSD.	256 GB NVMe SSD in a USB enclosure, including the fanxiang MD87 enclosure named in Part 1.
Cooling	Active cooling. A short tiny-prompt soak recorded CPU temperature between 58.7°C and 65.3°C with no throttle flag. Accelerator temperature was not available from that runtime. In the Part 6 session of 29 October the PWM fan ran at about 3,003 RPM at the start and about 9,500 RPM during generation. One sample reached 81.8°C and latched the historical soft-temperature-limit flag; active throttling bits were clear when checked.	A PWM cooling fan, fitted before the 29 October comparison. The fan ran at about 1,742 RPM at the start of the recorded session and about 5,300 RPM during generation. A mid-job sample read 70.8°C with no throttling flag.	Active cooling for the Pi, and the HAT heatsink on the accelerated machine.
Network	One machine on one local network until Part 6, which adds a second Pi on the same network. Heartbeats depended on the coordinator staying awake. Local inference on the Pi did not.	The same local network as Node 01. On 28 October it stopped answering on the local network while a model file was being copied. A later boot recorded an ext4 orphan-cleanup message, consistent with an unclean unmount. The cause was not established.	Gigabit Ethernet in preference to Wi-Fi, as a stable baseline rather than a measured result.
Power	Part 1 names an official 27 W USB-C supply as the intended supply for continuous operation. That is a design choice, not a measured energy result.	—	Official Raspberry Pi 27 W USB-C supply, named as the plan for continuous operation.