Raspberry Pi Launches AI HAT+ 2 With Hailo‑10H and 8GB RAM for On‑Device GenAI

Overview

• The $130 add‑on for Raspberry Pi 5 runs generative models locally using a Hailo‑10H NPU paired with 8GB of on‑board RAM to preserve privacy and reduce latency.
• Raspberry Pi lists launch models including DeepSeek‑R1‑Distill, Llama 3.2, Qwen2.5‑Coder 1.5B, Qwen2.5‑Instruct 1.5B, and Qwen2 1.5B, with more and larger options planned.
• The board delivers 40 TOPS (INT4) for LLM/VLM inference and offers computer‑vision performance broadly equivalent to the prior 26‑TOPS AI HAT, with tight integration into the Pi camera stack.
• Early benchmarks from Jeff Geerling report the Pi 5 CPU often outpacing the Hailo‑10H on LLMs, with the NPU closer on Qwen2.5‑Coder 1.5B and operating under a 3W power envelope versus roughly 10W for the Pi SoC.
• Hailo provides a hailo‑ollama backend and Dataflow Compiler with LoRA fine‑tuning support, though tests flagged outdated examples and errors when attempting concurrent vision and inference workloads.