Some links are affiliate links: if you buy through them we may earn a commission, at no extra cost to you. It never changes our assessments. Affiliate disclosure
The Raspberry Pi AI HAT+ 2 (Hailo-10H) adds a Hailo-10H neural accelerator with its own 8 GB of memory to a Raspberry Pi 5. Unlike the first AI HAT+ (vision only), it can run generative models. The question is which ones.
Yes, but only Hailo's models
The HAT does not run llama.cpp or arbitrary GGUF files. It runs models that Hailo has compiled for its chip, published in its GenAI model zoo. At the time of writing these are small, 1–2B-parameter models:
- Qwen3 1.7B, Qwen2.5 1.5B, Qwen2.5-Coder 1.5B, Qwen2 1.5B
- Llama 3.2 1B
- DeepSeek-R1-Distill-Qwen 1.5B
- Vision-language: Qwen2-VL 2B and Qwen3-VL 2B
- Whisper for speech recognition
Hailo provides hailo-ollama, a server with an Ollama-compatible API, so tools that speak to Ollama can use it.
What it is good for
- An offline voice assistant: Whisper plus a small Qwen model.
- Describing camera images with a 2B vision-language model next to classic vision AI.
- Simple classification, extraction and tool calls in a low-power project.
What it is not
It is not a general local-LLM accelerator. You cannot load an 8B model, your own fine-tune or a new release until Hailo compiles it. For a useful home chat assistant, even a 12 GB graphics card such as the NVIDIA GeForce RTX 3060 12 GB runs far better models.
The alternative at the edge
The NVIDIA Jetson Orin Nano Super Developer Kit costs more but runs any small model through CUDA, llama.cpp and Ollama, 3–4B models with moderate contexts in its 8 GB (shared with the system). If flexibility matters more than the Raspberry Pi ecosystem, choose the Jetson.