The arrival of the Raspberry Pi AI HAT+ 2 This represents a new step for those who want to work with artificial intelligence directly in a Raspberry Pi 5 without relying on the cloud. This expansion board combines a dedicated neural accelerator and its own memory to offload much of the AI ​​work from the main processor and enable more ambitious uses in the field of generative AI.
With a recommended price of around $130 (The final price in Spain and the rest of Europe will depend on official distributors.) The AI ​​HAT+ 2 is positioned as a relatively affordable option for experimenting with language models and computer vision at the edge. It's not a replacement for high-end solutions, but it is a striking piece for projects in... IoT, automation, prototyping, and training.
What exactly is the Raspberry Pi AI HAT+ 2
The Raspberry Pi AI HAT+ 2 is a expansion plate for the Raspberry Pi 5, which connects via the PCI Express interface integrated into the motherboard itself. It is the direct successor to the first AI HAT+, launched in 2024, which focused almost entirely on computer vision tasks thanks to the Hailo-8 and Hailo-8L accelerators.
In this new model, Raspberry Pi Holdings plc is betting on the neural network accelerator Hailo-10H and adds 8GB LPDDR4X memory dedicated on the card itself. This combination is designed to support workloads of Generative AI, especially large language models in reduced size and vision-language models, directly on the device.
By having DRAM memory Within the HAT itself, the system can separate the resources of the Raspberry Pi 5 (CPU, GPU, and main RAM) from those of the AI ​​accelerator. This allows the main board to handle application logic, communications, and user interface, while the Hailo-10H coprocessor takes over much of the inference.
Hardware, performance and differences with the first AI HAT+
The heart of the new accessory is the Hailo-10H NPU, a specialized neural network accelerator that, according to Raspberry Pi and Hailo, offers up to 40 TOPS inference performanceThe documentation mentions figures in INT4 and INT8This indicates that some of the peak performance is achieved through aggressive quantization, which is very common in edge AI deployments. Although other embedded platforms such as the Nvidia Jetson T5000 They offer alternatives with different architecture.
The consumption of Hailo-10H is limited to about 3 WThis helps keep the system within reasonable temperature and power limits, which is crucial for embedded projects and compact enclosures. However, this restriction means that raw performance won't always surpass that of the Raspberry Pi 5 working alone on certain tasks, especially when comparing workloads highly optimized for the integrated CPU and GPU.
Compared to the first AI HAT+, the main improvement is in the leap from the 13/26 TOPS of the Hailo‑8L and Hailo‑8 to the 40 TOPS, in addition to the appearance of 8 GB of RAM on board. The original model was more focused on object detection, pose estimation, or segmentation, while this second generation opens the door to LLMs and multimodal models, maintaining vision capabilities.
Raspberry Pi has emphasized that the integration with the camera environment It should be virtually transparent to existing projects, as integration with the company's software stack is maintained, avoiding the need to rewrite all the computer vision code from scratch.
Supported AI models and local LLM usage
One of the most striking points is the possibility of executing local language models based on the Raspberry Pi 5, leveraging the power of the NPU and the dedicated memory of the HAT. During the launch, the company mentioned a number of LLMs between 1.000 and 1.500 million parameters as a starting configuration.
Among the models mentioned are DeepSeek‑R1‑Distill, Llama 3.2, Qwen2, Qwen2.5‑Instruct and Qwen2.5‑CoderThese relatively compact models enable use cases such as basic chat, code generation, text translation, or scene description, all without sending data to external servers.
The plate is designed to take advantage of techniques of quantization and optimized inferenceThis allows models of this size to run with low latency and maintain low power consumption. Raspberry Pi and Hailo have demonstrated examples of language translation and responses to simple questions, all executed entirely on the device.
However, it's important to be clear that the AI ​​HAT+ 2 It is not designed for mass production models such as full versions of ChatGPT, Claude, or the larger Meta LLMs, whose number of parameters ranges in the hundreds of billions or even trillions. In these cases, the barrier is not only computing power, but above all, the memory required.
Memory limitations and comparison with a 16GB Raspberry Pi 5
While the 8 GB of dedicated LPDDR4X RAM represents a significant improvement over the first-generation AI HAT+, this figure has clear implications. Many medium-sized, context-rich, quantized language models can easily require more than 10 GB of memoryTherefore, for the time being, the accessory is geared towards smaller models or those with reduced screen sizes.
When compared directly to a 16GB Raspberry Pi 5, the motherboard with the highest memory configuration still offers more headroom for load large models entirely into RAMespecially if part of the memory is given up for other tasks and the system is dedicated almost exclusively to AI.
In practice, the combination of a Pi 5 and the AI ​​HAT+ 2 makes sense when you want to separate the functionsLet the Hailo-10H handle inference, while the mainboard maintains a lightweight desktop, web servers, automation logic, or additional services. This way, the system remains usable while running language or vision models.
For those who only want to chat with a simple local AI, translate texts, or test basic assistants, the AI ​​HAT+ 2's balance of power, consumption, and cost may be sufficient. However, for the necessary memory For large-scale models or very extensive contexts, it will still be more practical to use equipment with more memory or the cloud.
Computer vision and simultaneous model execution
One of the advantages of this new generation is that Computer vision capabilities are not being abandoned which the first AI HAT+ already offered. The Hailo-10H maintains very similar performance to the Hailo-8 when running object detection and tracking models, human posture estimation, or scene segmentation.
Raspberry Pi indicates that the AI ​​HAT+ 2 is capable of simultaneously run vision and language modelsThis makes it interesting for projects that combine cameras and text: for example, smart cameras that generate descriptions, surveillance systems that classify events, or devices that offer summaries of what is happening in a scene.
In practical terms, scenarios such as the use of model types are mentioned. YOLO For real-time object detection, with refresh rates that can reach around 30 frames per second depending on the model's complexity and resolution. The idea is that the HAT will handle this workload while the Raspberry Pi 5 manages storage, the interface, and sending alerts.
The software ecosystem is still maturing: although there are examples and official frameworks, the parallel execution of multiple models (Vision, language, multimodal) remains an evolving field. In any case, integration with the Raspberry Pi camera stack greatly simplifies the setup process for those already working with official camera modules.
Use cases in Spain and Europe: IoT, industry and prototypes
The combination of low power consumption, small size, and local AI execution fits well with many of the digitization trends that are being seen in Spain and other European countries. In industrial environments, where constant access to the cloud is not always guaranteed or is not desirable for confidentiality reasons, a This type of solution is an alternative to options such as integrating Nvidia into chips. It can be especially useful.
Among the most frequently repeated uses in the documentation are applications of industrial automation, process control and facilities managementVisual inspection systems on production lines, real-time anomaly detection, access monitoring, or people counting are some of the cases that could benefit from combining vision and lightweight language models.
In the realm of home and commercial IoT, the AI ​​HAT+ 2 can serve as a foundation for local assistants, smart panels that interpret sensor data, cameras that describe scenes, or video analytics devices that don't need to upload images to external servers, thus helping to comply with privacy regulations increasingly demanding in the European Union.
It is also an option to consider as development kit For companies and startups considering integrating the Hailo-10H chip into their own end products. Testing the solution on a Raspberry Pi 5 allows for validating performance, power consumption, and stability before embarking on custom hardware designs.
User profile: from makers to professional developers
The AI ​​HAT+ 2 targets several profiles simultaneously. On the one hand, it targets the community of makers and enthusiasts who are already familiar with Raspberry Pi and want to take it a step further with generative AI or advanced vision projects without investing in dedicated workstations or ongoing cloud fees.
On the other hand, it clearly points to professional developers and startups for those seeking an affordable testing environment for embedded AI. Compared to solutions with discrete GPUs or NPUs for industrial PCs, this HAT offers a compact form factor, low power consumption, and a more affordable cost, at the expense of the performance levels of much more expensive platforms.
In educational and technical training environments in Spain and Europe, it can become a useful tool for bringing the Technical formation to practice, allowing students to experiment with real models on relatively inexpensive hardware.
However, for users who only want to run language models locally with the maximum possible margin, a Raspberry Pi 5 with 16 GB of RAM without a HAT may still be a simpler option, provided they accept that the CPU and GPU of the motherboard They will be the ones that support all the inference.
Software integration and resources to get started
Raspberry Pi has emphasized that the AI ​​HAT+ 2 PCIe interface and the familiar Raspberry Pi 5 environment, reducing friction for those already familiar with the ecosystem. Communication with the HAT is via the PCIe interface, and dedicated drivers and libraries allow AI workloads to be routed to the Hailo-10H.
Hailo provides users with a repository on GitHub and a Developer Zone where code examples, pre-built models, tutorials, and frameworks are offered to help you leverage both generative AI and computer vision. This makes it easy to start prototyping without having to build the entire stack from scratch.
At the time of launch, several are already being advertised ready-to-install language modelswith the promise of expanding the catalog with larger or more refined variants for specific use cases. Furthermore, the possibility of using techniques such as LoRA (Low-Rank Adaptation) to adapt the models to specific tasks without having to completely retrain them.
As is often the case with these types of solutions, the actual experience will depend on the software maturity levelSome analysts point out that there is still room for improvement in stability, tools, and support for simultaneous execution of multiple models, but the trend points to increasingly polished integration within the Raspberry Pi ecosystem.
Price, availability and practical aspects
As for the price, the Raspberry Pi AI HAT+ 2 has been announced with a suggested retail price of $130. In Spain and other European countries, the final price will depend on the specific product or service. exchange rate, taxes, and each distributor's policyTherefore, slight variations are to be expected.
The board is compatible with all versions of Raspberry Pi 5, from the 1GB models up to the 16 GB of RAMand connects via the SBC's own PCIe interface. This eliminates the need for additional adapters and simplifies installation in cases or chassis designed for the standard form factor.
Included in the package is a optional heatsink for the Hailo-10H. Although the NPU is limited to about 3W, the usual recommendation is to install that heatsink, especially if you are going to run intensive workloads for long periods or demanding benchmarks, since the chip can reach high temperatures.
At the time of the presentation, some specialist stores reported that limited stockThis is a common occurrence whenever a new popular Raspberry Pi accessory is released. Therefore, those wanting to secure one in the near future will need to keep an eye on availability from authorized European distributors.
The Raspberry Pi AI HAT+ 2 is positioned as an intermediate solution between the cloud and large AI servers, designed for those who need local processingData privacy and contained costs. It offers a relatively affordable way to combine computer vision and lightweight language models at the edge, with room to grow as the software matures and the catalog of available models expands, provided that the power and memory limitations inherent in a low-power device are accepted.