Overview
The Hailo-10H is an accelerator built on Hailo's second-generation neural core, positioned to bring generative AI onto edge devices. It was among the earlier edge accelerators to bring LLM inference into scope while keeping power draw very low.
Its best-known form factor is the Raspberry Pi AI HAT+ 2, which is built on this chip and adds 8GB of LPDDR4X dedicated to holding models.
Key Features
- A compute-to-power ratio: 40 TOPS at INT4 and 20 TOPS at INT8 with 2.5W typical power draw. That ratio is the main selling point and suits battery-powered or sealed fanless devices.
- Direct DDR interface: A direct DDR interface means model size is not capped by on-chip cache, allowing larger models such as LLMs and VLMs.
- Industrial and automotive grades: The industrial grade covers -40°C to 85°C and the automotive grade -40°C to 105°C, opening up production lines and in-vehicle scenarios.
- Software stack and frameworks: Supports TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX, with x86 and ARM host architectures and Linux, Windows and Android as operating systems. The Dataflow Compiler handles model porting and quantisation.
- Raspberry Pi AI HAT+ 2: The official expansion board is priced at $130 with 8GB of on-board LPDDR4X as a separate memory pool, so model loading does not eat into the Pi's system memory. Launch support covers Llama 3.2 1B, DeepSeek-R1-Distill 1.5B and several Qwen 1.5B variants, so the starting scope is the 1B to 1.5B range.
- Single-tenant limitation: One chip runs one model at a time. Vision detection and an LLM summariser cannot run concurrently, and swapping models incurs roughly a 600ms stall. Scenarios needing vision and language together typically use two cards or pick a different platform.
Use Cases
- Raspberry Pi owners adding local LLM inference to hardware they already own
- Edge deployment of 1B to 3B class models for voice assistants, command parsing and summarisation
- Devices on battery or solar where the power budget stays under 15W
- Embedded projects that need LLM inference alongside GPIO, camera and I²C work
Pros
- An outstanding efficiency ratio at 40 TOPS INT4 and 2.5W typical power
- Raspberry Pi AI HAT+ 2 gives models their own memory pool instead of using system RAM
- Industrial and automotive temperature grades widen the addressable scenarios
- Connects over PCIe or USB, with both M.2 2242 and 2280 formats available
Pricing
Accelerators are sold as modules in Chip On Board, M.2 2242 and M.2 2280 formats. The Raspberry Pi AI HAT+ 2 is priced at $130.
Summary
The Hailo-10H brings edge generative AI down to a concrete price point: a Raspberry Pi plus the expansion board comes in under $300 and runs 1B to 3B class models. The trade-off is a ceiling on model size, with 7B class work needing a different platform, and an LLM toolchain that is younger than the vision stack, so pinning a known-good hailort version is advisable for production. For someone who already owns a Pi and wants a small model running locally, this is currently one of the cheaper credible paths.
Version History
- Hailo-10H production availability (2026): M.2 2242 and 2280 modules shipping, with industrial and automotive temperature grades
- Raspberry Pi AI HAT+ 2 (2026): Built on the Hailo-10H with 8GB of dedicated memory at $130