This week’s hardware roundup covers an operating-system-free AI experiment, an open-source IP KVM, a $20 networking board, and a Qualcomm compute module built for edge AI and multi-camera projects.

The most interesting hardware stories are not always about faster processors. This week’s standouts remove part of the software stack, push a microcontroller into a new role, or focus tightly on one job.

Running a local language model on a Raspberry Pi is nothing new. What makes NightRun unusual is that it removes the conventional operating system from the equation.

NightRun is an open-source Rust runtime that boots as a UEFI application. Instead of loading Linux first, the firmware starts NightRun directly. The runtime loads the model, draws its interface, reads keyboard input, and handles inference without a Linux kernel, userspace, browser, or network stack underneath it.

It is not pure bare-metal software: NightRun still relies on UEFI Boot Services for keyboard input, framebuffer access, memory allocation, and multicore startup. It is more accurately described as running without a conventional operating system.

The project supports 64-bit x86 UEFI computers booting from USB and the Raspberry Pi 5 booting from microSD. Current models include Llama 3.2 1B and 3B, Granite 4.1 3B, and Qwen3 4B, with memory requirements ranging from 4GB to 8GB.

On an 8GB Raspberry Pi 5, the developer reports Granite 4.1 3B generating about three tokens per second with the current NEON implementation—slow by desktop standards, but usable enough for a real demo.

NightRun also streams the model into RAM, verifies it, and then seals the boot media. Any later disk access causes a deliberate hard fault, ensuring that inference runs entirely from memory.

The project is experimental and aimed at developers comfortable with bootable media and firmware settings. It is not replacing Linux; its appeal is seeing how little software a local language model actually needs.

Your inbox is full. Slack is piling up. Client messages need a response yesterday. Typing thoughtful replies to all of it takes hours you don't have.

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Speak like you would to a colleague — tangents and all — and get polished output. Emails, Slack, LinkedIn, WhatsApp, whatever's open.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Try Wispr Flow free

The ESP32-P4 is also being pushed into a role normally handled by a Linux SBC or commercial appliance.

ESP-KVM is an open-source IP KVM built around the ESP32-P4 and a Toshiba TC358743 HDMI-to-MIPI CSI bridge. It captures HDMI output from a target computer, presents itself as a USB keyboard and mouse, and serves the video feed and controls through a browser.

Because it works below the operating-system level, ESP-KVM can reach a machine that will not boot normally. That makes it useful for entering BIOS or UEFI settings, selecting a boot device, reinstalling an operating system, or recovering a remote machine after a failed update.

It builds on Jonathan Rowny’s earlier p4kvm HDMI-capture proof of concept, adding a fuller web interface, authentication, encrypted access, input handling, and remote-management features.

Features include MJPEG and H.264 streaming, keyboard and mouse control, mobile input, virtual boot media, firmware rollback, Wake-on-LAN, optional ATX power control, MQTT, and WireGuard.

At 1080p, the developer measured about 20 frames per second with MJPEG and five to seven frames per second with H.264 on revision 1.3 silicon. H.264 uses less bandwidth but is slower because pre-3.0 ESP32-P4 chips require an extra color-space conversion before encoding. Revision 3.0 removes that step, although it had not yet been benchmarked.

ESP-KVM supports HTTPS and password authentication, but it has not undergone a formal security review. The developer recommends keeping it on a trusted network or accessing it through a VPN rather than exposing it directly to the internet.

ESP-KVM shows how much formerly expensive IP KVM functionality can now fit into a microcontroller-class device.

FriendlyElec’s NanoPi R28S is built around a short list of networking jobs rather than general-purpose flexibility.

The 45 × 45mm bare board starts at $20, while a bundle with a CNC-machined metal enclosure costs $29 before shipping or import fees.

It uses Rockchip’s RK3528A with four Cortex-A53 cores at up to 1.8GHz, a Mali-450 GPU, and hardware decoding for 4K H.264 and H.265 video. Every version includes 1GB of LPDDR4 memory, a microSD slot, and support for an optional eMMC module.

Networking is the main attraction: two Gigabit Ethernet ports, Wi-Fi 6, and Bluetooth 5.3. One USB-C connector provides 5V power and supports eMMC flashing, while a second is dedicated to serial debugging.

FriendlyElec lists Debian 13 Core, Ubuntu 24.04 Core, FriendlyWrt, Alpine Linux, OpenMediaVault, Buildroot, and Proxmox VE images on Linux 6.1 LTS. The fixed 1GB of RAM limits heavier server and virtualization workloads.

There is no display output, standard USB host port, or GPIO header. That makes the R28S less flexible than a normal SBC, but it also helps explain the low price and compact footprint.

For an OpenWrt router, travel firewall, VPN gateway, or network monitor, the trade-off makes sense. The R28S is cheap, compact, and already has the two Ethernet ports these projects often need.

Radxa’s CM-Q64, a 55 × 40mm compute module based on Qualcomm’s Dragonwing QCS6490 platform is tiny but mighty.

Its eight-core Kryo 670 CPU combines one core at up to 2.7GHz, three performance cores at up to 2.4GHz, and four efficiency cores at up to 1.9GHz. It also includes an Adreno 643 GPU, a Hexagon 770 AI engine rated for up to 12 TOPS, and configurations with up to 16GB of LPDDR5 memory.

Storage options include optional eMMC 5.1 or UFS 3.1, microSD expansion, and PCIe Gen 3 ×2 for NVMe. Depending on the module and carrier board, networking can include Gigabit or 2.5-Gigabit Ethernet, Wi-Fi 6, and Bluetooth 5.4, so buyers will need to check the exact configuration.

The multimedia hardware may be the bigger selling point. The CM-Q64 supports 4K60 decoding, 4K30 encoding, HDMI 2.0, DisplayPort 1.4 over USB-C, and MIPI DSI. Qualcomm’s Spectra 570L triple ISP can support up to four camera inputs, depending on lane configuration and software support.

That makes the module a strong fit for robotics, machine vision, smart cameras, industrial automation, and edge-AI systems that need to process several video feeds while running local inference.

Radxa currently lists Qualcomm Linux, Ubuntu, and Windows support. The company also says the module will remain available until at least July 2036, an important detail for long-lived embedded products.

Its value will depend on carrier boards, documentation, drivers, and access to the GPU, AI engine, cameras, and codecs. Even so, it packs substantial compute and multimedia hardware into a compact module.

These four releases show how broad the small-computing market has become. NightRun strips away the conventional operating system. ESP-KVM turns a RISC-V microcontroller into a remote-management appliance. The NanoPi R28S focuses on networking at the lowest practical price, while Radxa’s CM-Q64 targets demanding AI and embedded-vision products.

They serve different users and still have rough edges, but none is simply a familiar board with a faster processor. Each rethinks what a small computer should do—and that is usually where the best hardware stories begin