A Hailo-10H and its own memory
The AI HAT+ 2 connects to a Raspberry Pi 5 and adds a Hailo-10H neural accelerator rated by its makers at 40 TOPS for INT4 inference. Unlike the earlier AI HAT+, it includes 8GB of dedicated onboard memory for models and runtime data.
Raspberry Pi launched the board at $130 and described it as available immediately on January 15. Price and stock are time-sensitive, and the headline TOPS figure does not reveal application-level latency, accuracy, power draw or sustained performance.
Small models by design
Launch examples included 1-billion to 1.5-billion-parameter versions of Llama 3.2, DeepSeek-R1-Distill and Qwen. Raspberry Pi says edge models for this class of hardware typically fall between 1 and 7 billion parameters.
The company explicitly says those models are not designed to match the knowledge of much larger cloud systems. Their value is narrower: local processing, constrained tasks, predictable deployment and the possibility of operating without a network connection.
Compatibility still depends on compiled models
The board supports language models, vision-language models and conventional computer-vision workloads. Raspberry Pi says it remains integrated with libcamera, rpicam-apps and Picamera2, but applications need model files compiled for the Hailo-10H rather than the earlier accelerator.
Task-specific Low-Rank Adaptation is available through Hailo's Dataflow Compiler. That can tune a small base model for a narrow application; it does not create missing knowledge or guarantee safe, reliable output.
This is not a performance review
LUMACTA NEWS has not tested tokens per second, first-token latency, camera throughput, power use, heat, installation reliability or output quality. A buying recommendation would require those measurements across representative workloads.
The hero is a real photograph of the exact retail package. Its printed words and board illustration are native packaging, not an artificial news cover or an editorial text overlay.
Sources & Methods
LUMACTA NEWS separated rated hardware specifications from untested application performance and used an independently photographed retail package under a traceable Creative Commons licence.
- Raspberry Pi — Introducing the AI HAT+ 2 — manufacturer announcement
- Wikimedia Commons — Raspberry Pi AI HAT+ 2 box — image source and licence record
- Creative Commons — CC BY-SA 3.0 — image rights source
