Hailo

What is Hailo?

Hailo is an Israeli semiconductor company that designs processors for running AI models on edge devices. Its accelerators handle inference locally, so cameras and robots can analyse video offline.

Founded in 2017 by Orr Danon and Avi Baum, Hailo ships the Hailo-8, Hailo-8L, Hailo-10H and Hailo-15 families. Microchip Technology completed its acquisition of Hailo on 21 September 2026 and says it will keep supporting the existing products and software.

Key Features

• Hailo-8 accelerator: Hailo-8 delivers 26 TOPS at about 2.5W typical power. It keeps the memory it needs on the chip instead of relying on external DRAM, which helps efficiency in small devices.
• Hailo-10H for generative AI: Hailo-10H adds a direct LPDDR4/4X memory interface and reaches 40 TOPS at INT4. That memory lets LLMs, VLMs and Stable Diffusion run on the device without a cloud connection.
• Hailo-15 vision processors: Hailo-15 chips combine AI acceleration with image processing for smart cameras. Camera makers can build analytics into the camera itself rather than adding a separate accelerator.
• Dataflow Compiler: The compiler converts and optimises trained TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX models for Hailo chips. Engineers can keep their existing training pipeline and only change the final deployment step.
• HailoRT and Model Zoo: HailoRT runs models on the device, and the Model Zoo supplies pre-trained networks. TAPPAS and an examples repository add ready demos for detection, pose estimation and segmentation.
• Several form factors: Hailo-8 ships as a bare chip, as M.2 and mPCIe modules and as a PCIe card. Hardware designers can pick the format that fits their board and enclosure.
• Multi-stream processing: Hailo-8 handles several video streams and models at once, and multiple units can be combined for more capacity. This suits camera hubs and industrial inspection lines.
• Industrial and automotive grades: Industrial parts run from -40°C to 85°C, and automotive parts up to 105°C. Hailo-8 is AEC-Q100 Grade 2 compliant and meets ISO 26262 ASIL-B requirements for vehicle use.

Pros

• High efficiency: Hailo-8 reaches 26 TOPS at about 2.5W, and The Register reported the 8L draws roughly one watt per three TOPS. That suits fanless and battery-powered devices.
• Local processing: Inference happens on the device, so video and sensor data need not leave the camera. This reduces cloud costs, latency and privacy exposure, although the host device still needs its own security.
• Affordable entry: The Raspberry Pi AI Kit launched at a reported $70, and the AI HAT+ 2 costs $200. Hobbyists and students can test Hailo chips cheaply before any volume order.
• Framework support: Models from TensorFlow, PyTorch, Keras and ONNX compile for Hailo chips. Hosts can run Linux, Windows or Android on x86 or ARM processors, which covers most embedded and desktop boards.
• Offline generative AI: Hailo-10H runs small language and vision-language models without a network connection. Products can offer chat or image captioning features in places with poor connectivity without depending on cloud services.

Cons

• Not a software app: Hailo sells chips and modules, so you need hardware and embedded engineering skills. Buyers looking for a ready AI service will find nothing to log into.
• Quote-based pricing: Hailo does not publish chip or module prices. Volume buyers must request quotes through distributors, which slows early budgeting, although Raspberry Pi kits offer a cheap way to prototype.
• Compiler licence: One analysis flags that the Dataflow Compiler comes under a revocable licence. Teams building long-life products should review those terms with legal advisers before committing to the toolchain.
• Limited model size: The Hailo-8L kit offers 13 TOPS, well below some rival NPUs, and very large models such as Llama 3 70B still need the cloud. Hailo chips suit small, efficient models.
• Ownership change: Microchip completed its acquisition of Hailo on 21 September 2026. Microchip says support for existing products continues, but roadmaps, pricing and distribution channels may change over time.

Who is Using Hailo?

• Camera makers: Smart security and retail analytics devices.
• Automotive suppliers: Driver assistance and in-cabin systems.
• Robotics teams: On-board vision for robots and drones.
• Hobbyists and students: Raspberry Pi AI projects at home.

Pricing

Hailo sells chips and modules by quote, so prices depend on volume and distributor (checked October 2026).

Kit prices come from Raspberry Pi and press reports, so confirm current figures with a stockist.

What Makes Hailo Unique?

Hailo-8 keeps the memory it needs on the chip itself rather than relying on external DRAM. That helps it reach 26 TOPS at around 2.5W. The Hailo-10H adds DDR so small language models run offline.

Compatibilities and Integrations

• Platforms: Linux, Windows and Android hosts on x86 or ARM.
• Frameworks: TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX.
• Hardware partners: Raspberry Pi 5 kits and boards from other makers.
• Software: Dataflow Compiler, HailoRT, TAPPAS and the Model Zoo.

Hailo Tutorials

A Raspberry Pi kit is the cheapest way to try a Hailo chip.

1. Fit the Hailo module or HAT to a Raspberry Pi 5.
2. Install the Hailo packages and runtime on the Pi.
3. Run a demo from the hailo-rpi5-examples repository.
4. Pick a pre-trained network from the Model Zoo.
5. Compile your own model with the Dataflow Compiler.

How We Rated It

Summary

Hailo suits engineers and makers who need fast, low-power AI inference on cameras, robots or single-board computers. Its main limitation is that it is hardware, and compiling models takes skill. Start with a Raspberry Pi kit to test your model before ordering modules.