Overview
- The $130 add‑on is available now for Raspberry Pi 5 and launches with Llama 3.2, DeepSeek‑R1‑Distill, and Qwen 2/2.5 variants at roughly 1.5 billion parameters.
- The board delivers a Hailo‑10H accelerator rated at 40 TOPS (INT4) with 8GB onboard RAM, connects over PCIe, integrates with Raspberry Pi’s camera stack, and ships with a recommended heatsink.
- Independent testing found the Pi 5’s CPU generally outperformed the Hailo 10H on supported LLMs, with the NPU constrained to about 3W versus up to 10W for the Pi SoC.
- Reviewers reported smooth single‑model demos using hailo‑ollama, yet noted early software rough edges and instability when attempting simultaneous vision and LLM workloads.
- Coverage positions the HAT+ 2 for niche or development use where private, low‑latency inference is needed, as it cannot match larger cloud models and may be less flexible than a 16GB Pi for bigger local LLMs.