- The Raspberry Pi AI HAT+ 2 incorporates a Hailo-10H NPU with up to 40 TOPS and 8 GB of dedicated RAM.
- It allows you to run lightweight language models and computer vision locally, without depending on the cloud.
- It maintains compatibility with Raspberry Pi 5 and its camera ecosystem, but is limited to compact LLMs.
- Its price is around $130 and it targets IoT, industry, education and prototyping projects in Europe.
The arrival of the Raspberry Pi AI HAT+ 2 This marks a new step for those who want to work with artificial intelligence directly in a Raspberry Pi 5 without permanently relying on the cloud. This expansion board adds a dedicated neural accelerator and its own memory, so that much of the AI processing is moved off the main CPU, allowing for more ambitious generative AI and computer vision projects.
With a recommended price of around $130 (The final price in Spain and the rest of Europe will vary depending on taxes and official distributor margins.) The AI HAT+ 2 positions itself as a relatively affordable option within the embedded AI ecosystem. It doesn't compete with large servers or dedicated GPUs, but it does offer an interesting balance between cost, power consumption, and performance. IoT, automation, education, and prototyping.
What is the Raspberry Pi AI HAT+ 2 and how does it differ from the first generation?
The Raspberry Pi AI HAT+ 2 is a official extension plate Designed for the Raspberry Pi 5, it connects via the motherboard's integrated PCI Express interface and also uses the GPIO connector for mounting. It's the direct successor to the first AI HAT+, released in 2024, which was offered in variants with accelerators. Hailo‑8L (13 TOPS) and Hailo‑8 (26 TOPS) and was very focused on computer vision tasks.
In this second generation, Raspberry Pi is betting on a Hailo-10H neural network accelerator accompanied by 8 GB of LPDDR4X memory dedicated on the card itself. This combination is designed to support workloads of generative AI at the edge, such as compact language models, vision-language models, and multimodal applications that combine image and text.
The fact of incorporating integrated DRAM This means that running AI models doesn't directly consume the Raspberry Pi 5's main memory. The motherboard can focus on application logic, the user interface, connectivity, or storage, while the NPU handles the bulk of the inference. In practice, this helps keep the system usable while AI models run in the background.
According to Raspberry Pi itself, the transition from the first AI HAT+ to this new model is virtually transparent For projects that already used Hailo-8 accelerators, the integration with the company's camera environment and software stack is maintained, avoiding massive rewrites.
Hardware, performance and power consumption: up to 40 TOPS with the Hailo-10H NPU
The heart of the AI HAT+ 2 is the Hailo-10HA specialized neural network accelerator designed to efficiently run inferences on low-power devices. Raspberry Pi and Hailo are talking about up to 40 TOPS of performance (teraoperations per second), figures obtained with quantization in INT4 and INT8, very common when models are deployed at the edge.
One of the key points is that the chip is limited to a power of around 3W power consumptionThis allows it to be integrated into compact enclosures and embedded projects without significantly increasing cooling requirements or electricity bills, which is important for devices that may be active 24/7. However, this restriction means that the gross yield It will not always be superior to what the Raspberry Pi 5 itself can offer when its CPU and GPU are pushed to their limits in certain highly optimized workloads.
Compared to the previous model, the leap is clear: it goes from 13/26 TOPS with Hailo‑8L/Hailo‑8 It achieves 40 TOPS with Hailo-10H, and for the first time, 8 GB of dedicated onboard memory is added. The first AI HAT+ excelled at tasks such as object detection, pose estimation, and scene segmentation; the new version maintains these types of applications but broadens its focus to language models and multimodal uses.
Even so, Raspberry Pi itself clarifies that, in certain vision operations, the practical performance of the Hailo-10H may be similar to the 26 TOPS of the Hailo-8, due to the way the workload is distributed and the architectural differences. The major improvement, more than in raw computer vision power, lies in the possibilities it opens up for LLM and local generative models.
The plate comes with a optional heatsink for the NPU. Although power consumption is limited, the usual recommendation is to install it, especially if you are going to run intensive AI tasks for a long time or demanding performance tests, to prevent the chip from reducing frequencies due to temperature.
Supported language models and local LLM usage
One of the most striking aspects of the AI HAT+ 2 is its ability to run language models locally on a Raspberry Pi 5, without sending data to external servers. During the presentation, Raspberry Pi and Hailo highlighted a range of models, including 1.000 and 1.500 million parameters as a starting point.
Among the compatible LLMs offered at launch are DeepSeek‑R1‑Distill, Llama 3.2, Qwen2, Qwen2.5‑Instruct and Qwen2.5‑CoderThey are relatively compact models, designed for tasks such as basic chat, text writing and correction, code generation, simple translations, or scene descriptions from image and text inputs.
The initial tests shown by the company include examples of translation between languages and answers to simple questions executed entirely on the Raspberry Pi 5 supported by the AI HAT+ 2, with low latency and without significantly impacting the overall system usability. Processing is performed on the Hailo-10H coprocessor and does not require connecting the device to the cloud.
It should be made clear that this solution is not intended for mass-market models such as full versions of ChatGPT, Claude, or the larger LLMs at Metawhose sizes are measured in hundreds of billions or even trillions of parameters. In those cases, the problem is not only computing power, but above all the memory required to host the model and its contexts.
Raspberry Pi itself insists that users should be aware that they are working with smaller models trained on more limited datasetsTo compensate for this restriction, the focus is placed on techniques such as LoRA (Low-Rank Adaptation)which allow models to be adjusted to specific use cases without the need to completely retrain them, adding lightweight adaptation layers on top of the existing base.
Memory, limitations and comparison with a 16GB Raspberry Pi 5
The inclusion of 8 GB of dedicated LPDDR4X RAM This is one of the major new features of AI HAT+ 2, but it also clearly defines the types of models that can be run. Many medium-sized quantized LLMs, especially if you want to handle a broad context, may easily need more than 10 GB of memoryTherefore, the accessory is geared towards lightweight models or those with tighter context windows.
If you compare it to a Raspberry Pi 5 16GB Even without HAT, motherboards with more memory still have an advantage when loading relatively large models directly into RAM, provided a significant portion of that memory is dedicated exclusively to AI and other tasks are sacrificed. In that scenario, the integrated CPU and GPU handle all the inference, resulting in increased workload.
The AI HAT+ 2 proposal makes more sense when looking for separate responsibilitiesLet the Hailo-10H NPU handle the AI calculations and free up the Raspberry Pi 5 to maintain a lightweight desktop environment, web services, databases, automations, or the presentation layer of an application.
For those who only want to have one local assistant Relatively simple and capable of chatting, translating texts, or assisting with minor programming tasks without sending data to third parties, the AI HAT+ 2's balance of power, consumption, and cost may prove sufficient. However, for projects requiring large models or extremely extensive contexts, using devices with more memory or cloud infrastructure will remain more practical.
Another point to consider is that, although the HAT's 8 GB helps offload memory, the version of 16 GB of the Raspberry Pi 5 It still outperforms the add-in board in total capacity, so in certain RAM-intensive workflows that configuration will continue to be preferable.
Computer vision and simultaneous model execution
The AI HAT+ 2 does not abandon the feature that made the first generation popular: the computer vision applicationsThe Hailo-10H is capable of running object detection and tracking models, human pose estimation, or scene segmentation with performance that, in practice, remains in line with what the Hailo-8 offered at 26 TOPS.
Raspberry Pi indicates that the new board can simultaneously run vision and language modelsThis makes it attractive for projects where the camera and text processing need to work together. For example, surveillance systems that classify events and generate descriptions, smart cameras that explain what is happening in a scene, or devices that combine visual recognition with report generation.
In specific scenarios, family models are mentioned. YOLO For real-time object detection, with refresh rates that can reach around 30 frames per second depending on the resolution and complexity of the model. The idea is that the NPU will handle this task while the Raspberry Pi 5 manages storage, network, notifications, and display.
The software ecosystem surrounding AI on Raspberry Pi is still maturing. Although a collection of examples, frameworks and tools For both Raspberry Pi and Hailo, parallel execution of multiple models (vision, language, multimodal) continues to be an evolving field and may require fine-tuning in each project.
In any case, integration with the official Raspberry Pi camera stack This simplifies life for those already working with the brand's camera modules. The AI HAT+ 2 integrates directly with that environment, so many existing vision projects can migrate to the new board with relatively minor changes.
Use cases in Spain and Europe: industry, IoT and educational projects
The combination of low power consumption, small size and local AI processing This aligns well with the digitalization trends being implemented in Spain and other European countries. In industrial sectors where stable cloud access is not always guaranteed or where strict confidentiality requirements exist, a solution of this type can be particularly attractive.
Among the most frequently used terms in official documentation are projects for industrial automation, process control and facilities managementVisual inspection systems on production lines, real-time anomaly detection, access control, or counting people in buildings are examples where the combination of vision and lightweight language models can add value without the need to deploy much more expensive AI infrastructures.
In the field of Home and business IoTThe AI HAT+ 2 can serve as a foundation for local assistants running on a Raspberry Pi 5, dashboards that interpret sensor data, cameras that describe scenes, or devices that analyze video without uploading images to external servers. This approach helps comply with increasingly stringent data protection regulations in the European Union.
It can also be an interesting tool as development kit For European companies and startups considering integrating the Hailo-10H chip into end products. Testing performance and stability on a Raspberry Pi allows for validating concepts before investing in custom hardware designs.
In the educational field, vocational training centers, universities, and specialized academies in Spain could use AI HAT+ 2 as a practice platform, bringing the Embedded AI and generative AI to students on accessible and relatively inexpensive hardware compared to other more expensive systems.
User profile and type of projects targeted
The Raspberry Pi AI HAT+ 2 targets several profiles. On the one hand, the broad community of makers and enthusiasts who already use the Raspberry Pi 5 and want to incorporate generative AI or advanced vision into their projects without making the leap to workstations with dedicated GPUs or depending entirely on cloud services.
On the other hand, he tries to seduce professional developers and startups that need a testing platform for embedded AI. Compared to solutions with eGPUs or NPUs integrated into industrial PCs, this board offers a compact form factor, very low power consumption, and a lower overall cost, although with a lower performance ceiling than much more expensive platforms.
For those already experienced with the first AI HAT+, the transition seems relatively simple: integration with existing software And the camera stack has been carefully designed to minimize the necessary changes. This is relevant for projects already underway that want to take advantage of the performance increase without rewriting everything.
At the other extreme, users who are only looking to run language models locally with the maximum possible memory margin may still find a Raspberry Pi 5 16GB Without HAT, assuming that the integrated CPU and GPU will handle all inference and that power consumption will be somewhat higher.
In short, the accessory seems to be carving out a niche as an intermediate solution: more powerful and flexible than a Raspberry Pi 5 working alone on certain AI tasks, but far from the performance of servers or dedicated GPUs, and with a focus on low power consumption, privacy and cost containment.
Hailo software integration, resources, and support
From a software perspective, Raspberry Pi has aimed to simplify the setup process as much as possible. The AI HAT+ 2 connects via the PCIe interface of the Raspberry Pi 5 and is natively recognized by the official operating system, allowing AI applications to run without overly complex setup steps for those already familiar with the environment.
Hailo provides users with a repository on GitHub and a Developer Zone It includes code examples, pre-configured models, tutorials, and frameworks designed for both generative AI and computer vision. It also includes tools for managing quantization, loading third-party models, and optimizing specific workflows.
At launch, the company has made available several ready-to-install language modelswith the promise of expanding the catalog with larger variants or those adapted to very specific use cases. Furthermore, it encourages the use of techniques like LoRa to adjust the models to the needs of each project without having to train them from scratch on enormous datasets.
As is often the case with these types of solutions, the actual experience will depend on the maturity level of the software ecosystemSome analysts point out that there is still room for improvement in tools, stability, and support for simultaneous execution of multiple models, but the trend in the Raspberry Pi ecosystem is moving towards increasingly polished integration.
In any case, to develop projects in Spain or other European countries, having official documentation, practical examples and an active community considerably reduces the barrier to entry for experimenting with embedded and generative AI in low-cost devices.
Price, availability and practical aspects in Spain and Europe
The Raspberry Pi AI HAT+ 2 has launched with a reference price of $130In Spain and the rest of Europe, the final amount will depend on the exchange rate, taxes, and each distributor's policyTherefore, it is expected that there will be small differences between stores and countries.
The motherboard is compatible with the entire line of Raspberry Pi 5From models with 1GB of RAM to versions with 16GB, the compatible Raspberry Pi is mounted using the familiar HAT format: it screws onto the board and connects via the GPIO header and the PCIe interface. Previous Raspberry Pi models that lack this interface are therefore excluded from the compatibility list.
In the initial stages following the announcement, some specialist distributors reported that Limited stockThis is now common practice with official Raspberry Pi hardware releases. Those wanting to secure a unit in the short term will need to keep an eye on availability from authorized European distributors and potential waiting lists.
In addition to the hardware, the purchase includes access to technical documentation and software resources for Raspberry Pi and Hailo, including GitHub examples, step-by-step guides, and materials for those new to embedded AI. This makes it easy for both individual users and small businesses to start experimenting without needing to invest in additional development tools.
In the European context, where the data privacy And as energy efficiency becomes increasingly relevant, the AI HAT+ 2 is presented as a piece that allows process sensitive information locally reducing dependence on remote data centers, which may be attractive to administrations, SMEs and independent developers looking for more controlled AI solutions.
The Raspberry Pi AI HAT+ 2 positions itself as an intermediate solution between the cloud and large AI servers: it offers a reasonably accessible way to combine computer vision and lightweight language models in a single device, keeping power consumption low and respecting privacy, but requiring in return that projects be designed within the limits of power and memory typical of hardware designed for low power consumption and low cost.
I am a technology enthusiast who has turned his "geek" interests into a profession. I have spent more than 10 years of my life using cutting-edge technology and tinkering with all kinds of programs out of pure curiosity. Now I have specialized in computer technology and video games. This is because for more than 5 years I have been writing for various websites on technology and video games, creating articles that seek to give you the information you need in a language that is understandable to everyone.
If you have any questions, my knowledge ranges from everything related to the Windows operating system as well as Android for mobile phones. And my commitment is to you, I am always willing to spend a few minutes and help you resolve any questions you may have in this internet world.