RASPBERRY PI AI HAT+ 2

Now available: Raspberry Pi launches the global release of its new Raspberry Pi AI HAT+ 2 add-on board, specifically designed to accelerate both vision models and selected Large Language Models (LLMs) as well as Vision-Language Models (VLMs). 

Many industrial edge programs are now required to deliver “AI capability” without taking open-ended platform risks. The new Raspberry Pi AI HAT+ allows for defining auditable platform requirements: a fixed host baseline, a defined acceleration scope, and explicit procurement parameters (SKU, MOQ, ordering multiple) that directly impact series planning and service commitments. 

About the new Raspberry Pi AI HAT+ 

The Raspberry Pi AI HAT+ 2 delivers performance similar to the 26 TOPS AI HAT+, while also providing acceleration for selected LLMs and VLMs. For OEMs, this means that vision-based designs can remain within an established performance class, while specific language functions or multimodal features can be evaluated and tested within a defined scope. In projects, this improves requirements’ traceability and reduces the risk of unbounded feature claims in specifications. It secures the decision of which AI features are productized at launch and which are deferred. 

 
The AI HAT+ 2 connects via PCI Express and is exclusively compatible with Raspberry Pi 5. For OEMs, this sets a clear platform boundary: hardware architecture, enclosure integration, and qualification planning must align with a Raspberry Pi 5-based design. In day-to-day projects, this reduces interface ambiguity and avoids later redesigns caused by incorrect host assumptions. It secures the decision whether the roadmap should standardize Raspberry Pi 5 as the host standard for the intended lifecycle. 

 
The board is based on the Hailo 10H Neural Network Accelerator and includes 8GB of LPDDR4X-4267 SDRAM onboard. This defines a stable accelerator subsystem with a specified local memory configuration, which is important for predictable verification and performance characterization on supported workloads. In projects, this supports the definition of a controlled test plan and clearer responsibility boundaries between host resources and accelerator resources. It supports the decision on validation depth and performance risk for the AI pipeline in series. 

 
The expected support at launch includes Llama-3.2-3B-Instruct and QWEN2.5-VL-3B, with ASR support (whisper-base). The specification of “LLM/VLM capable” turns into a concrete starting point that can be validated and documented. This enables the planning of a clearly defined test matrix, regression planning, and update governance for series systems, securing the decision on the initial release baseline and maintenance scope for supported models. 

Key Benefits 

• Hailo 10H Neural Network Accelerator with 8GB LPDDR4X-4267 SDRAM onboard: defines a stable accelerator subsystem and supports a controlled verification scope 
• PCIe attachment with exclusive Raspberry Pi 5 compatibility: sets the host baseline early and reduces architectural ambiguity 
• Similar performance for vision models as the 26 TOPS AI HAT+ 
• Defined model scope (vision plus acceleration of selected LLM/VLM/ASR): enables testable requirements and bounded feature commitments
• Expected LLM/VLM support at launch 

o Llama-3.2-3B-Instruct 
o QWEN2.5-VL-3B 
o ASR (whisper-base)