TL;DR
Raspberry Pi released the AI HAT+ 2: a $130 add-on with a Hailo 10H NPU and 8 GB of LPDDR4X aimed at running inference locally. Benchmarks show the Pi 5's CPU often outperforms the 10H on LLM workloads, while the HAT still offers strong machine-vision performance but faces software maturity issues.
What happened
Raspberry Pi launched the AI HAT+ 2, a $130 board combining a Hailo 10H neural processor and 8 GB of LPDDR4X memory. The Hailo chip advertises up to 40 TOPS INT8 inference at a 3W power envelope and is intended to run models standalone so the host Pi’s CPU and system RAM are less taxed. Tests run on an 8 GB Raspberry Pi 5 showed the Pi’s built-in CPU frequently outpacing the Hailo 10H for LLM tasks; the NPU was only competitive on a few models such as Qwen2.5 Coder 1.5B. Vision workloads ran significantly faster on the Hailo board—around 10x faster than the Pi’s CPU—but similar functionality is available from earlier Raspberry Pi AI accessories at lower cost. Attempts to run mixed workloads (vision and an LLM simultaneously) encountered segmentation faults and device errors, and some Hailo example packages were not yet updated to support the new 10H.
Why it matters
- Adds a local AI coprocessor option for Pi users who need on-device inference without consuming host RAM.
- Lower power draw and smaller size than eGPU approaches, making it potentially useful for power-constrained edge devices.
- Limits in RAM (8 GB) and NPU power (3W) restrict usefulness for medium-to-large LLMs compared with higher-RAM Pi configurations.
- Software and driver maturity will determine whether the mixed vision+inference promise is practical for developers and deployments.
Key facts
- Price: $130 for the AI HAT+ 2.
- Hardware: Hailo 10H NPU paired with 8 GB LPDDR4X RAM.
- Performance: Hailo 10H advertises 40 TOPS INT8 inference; earlier Hailo 8 delivered ~26 TOPS INT4 for vision.
- Power: Hailo 10H tops out at around 3W; Pi SoC power limits tested were up to ~10W.
- Test setup: Benchmarks were run on an 8 GB Raspberry Pi 5 to compare Pi CPU vs Hailo NPU on the same RAM footprint.
- LLM results: The Pi 5’s CPU often outperformed the Hailo 10H on LLM inference; Hailo was closest on Qwen2.5 Coder 1.5B.
- Vision results: Camera-based models ran about 10x faster on the Hailo board than on the Pi’s CPU.
- Compatibility issues: Hailo’s example packages were not yet updated for the 10H and attempts to run simultaneous vision+inference produced segmentation faults or 'device not ready' errors.
- Alternatives: Raspberry Pi previously offered an AI HAT (~$110) and an AI Camera (~$70) for vision tasks.
What to watch next
- Updates from Hailo and Raspberry Pi to fix driver/example support and to resolve segmentation faults when running two models simultaneously (not confirmed in the source).
- Whether software and model-optimization tools emerge to make mixed vision+LLM workloads stable and efficient on the 10H (not confirmed in the source).
- Adoption of the 10H in commercial edge devices (self-checkout, embedded vision) versus remaining primarily a development kit (not confirmed in the source).
Quick glossary
- NPU: Neural Processing Unit: a specialized chip designed to accelerate machine learning inference workloads.
- LPDDR4X: A low-power variant of DDR4 DRAM commonly used in mobile and embedded devices for system and graphics memory.
- TOPS: Tera Operations Per Second: a rough measure of an accelerator’s raw throughput for integer operations during inference.
- Quantization: A technique that reduces model size and compute by using lower-precision numeric formats, often trading some accuracy for efficiency.
Reader FAQ
Does the AI HAT+ 2 increase the Raspberry Pi’s system RAM?
No — the HAT includes 8 GB of LPDDR4X for the coprocessor, but it does not change or upgrade the Pi’s onboard system RAM.
Is the Hailo 10H faster than the Pi 5 CPU for LLMs?
In the tests cited, the Pi 5’s CPU frequently outperformed the Hailo 10H on LLM inference; the 10H was competitive in only a few models.
Is the AI HAT+ 2 a good choice for vision processing?
Yes—vision models ran much faster on the Hailo board (around 10x faster than the Pi CPU), but similar capabilities exist in earlier, less expensive Raspberry Pi AI accessories.
Can the HAT run vision and LLM models at the same time?
Attempts to run dual workloads encountered segmentation faults and device errors in the cited tests; resolution depends on software updates and is not yet available.
Raspberry Pi's new AI HAT adds 8GB of RAM for local LLMs Jan 15, 2026 Today Raspberry Pi launched their new $130 AI HAT+ 2 which includes a Hailo 10H…
Sources
- Raspberry Pi's New AI Hat Adds 8GB of RAM for Local LLMs
- Raspberry Pi AI HAT+ 2 adds 8 GB RAM next to Hailo-10H
- Buy a Raspberry Pi AI Kit
- Raspberry Pi AI+ HAT disappointment – Hailo Community
Related posts
- Opera One R3 arrives with color-coded Tab Islands and new soundscapes
- Eigent: Open-source Cowork Desktop for Multi-Agent AI Workforces
- Zhipu AI says GLM-Image was trained solely on Huawei Ascend and Kunpeng