Run a 7-billion-parameter language model on a passively cooled box you can bolt to a DIN rail. That is the pitch behind the MYiR MAC-B5760, a fanless industrial PC built around the Rockchip RK3576 and aimed squarely at on-device AI work. The base configuration ships with 8GB of LPDDR5 and 64GB of eMMC flash, and the higher tier adds an M.2 accelerator card that turns the unit into a local LLM and vision box without a cloud connection.

The RK3576 itself is a familiar quantity for anyone following the open hardware scene: four Cortex-A72 cores at 2.2GHz, four Cortex-A53 cores at 1.8GHz, a Mali-G52 MC3 GPU with OpenGL ES 3.2 and Vulkan 1.1, and a 6 TOPS NPU that handles INT4 through TF32 mixed operations. The video block decodes H.264, H.265, VP9, AV1, and AVS2 up to 8K at 30fps. Connectivity is generous for the class, with two Gigabit Ethernet ports, four USB 3.0 ports, HDMI 2.1 up to 4Kp120, a Mini DisplayPort up to 4Kp60, and a USB-C port wired for serial console access. An M.2 Key-M 2280 slot takes either an NVMe SSD or an AI accelerator, while a mini PCIe socket and micro SIM slot add optional 4G LTE or 5G, alongside an optional WiFi 6 and Bluetooth 5.4 module.

The software story is where this gets interesting for the self-hosting crowd. MYiR provides a Linux BSP on the Rockchip 6.1 vendor kernel with open-source drivers, plus prebuilt Debian and Yocto images, including a Yocto build with a Preempt RT kernel for real-time work. The mainline picture has also improved sharply over the past year: RK3576 support landed in Linux 6.12 covering clocks, power management, eMMC, networking, and the GPU, and Collabora's H.264 and H.265 video decoder support for the RK3576 and RK3588 was merged into Linux 7.0, narrowing the gap with the vendor tree. The open-source rkllama project builds on top of Rockchip's RKLLM runtime to expose an Ollama-compatible API server for RK3576 and RK3588 hardware, implementing the standard /api/chat and /api/generate endpoints so Open WebUI and other Ollama-compatible clients connect without custom integration work.

The optional RM1828MC0-F M.2 card is the headline upgrade. Based on the Rockchip RK1828, it adds 20 TOPS and 5GB of dedicated memory for computer vision and language or vision-language models up to 7 billion parameters. The chip pairs with Rockchip's RKLLM toolkit, which converts Hugging Face models such as LLaMA, Qwen2, and Phi-2 into NPU-accelerated formats, with reported throughput north of 2000 tokens per second on prefill and around 120 tokens per second on decode for the RK182X family.

The chassis measures 150 x 101 x 47 mm (5.9 x 4.0 x 1.9 inches) without the mounting bracket and weighs 575 grams (1.3 lbs), runs from a 12V/5A barrel jack, and is rated for 0 to 70°C (32 to 158°F). Two SKUs are on sale now: the standard MAC-B5760-CC810000 with 8GB RAM and 64GB flash starts at $300 (€280), while the MAC-B5760-CC81000C bundles the RK1828 AI module for $600 (€550). Both are detailed on the MYiR product page. It joins a growing field of fanless RK3576 edge machines that includes the Mekotronics R57, the Firefly EC-R3576PC-FD, and the reComputer RK3576.