US What the Hailo-10H is
The Hailo-10H is the first Hailo accelerator built for generative AI. Hailo, an Israeli chip designer whose acquisition by Microchip completed on 21 September 2026, is best known for the Hailo-8 vision accelerator. The Hailo-10H keeps the same dataflow architecture but adds a memory interface, so it can run models far too large to fit on the chip itself.
The Hailo-8 keeps a vision model entirely on the chip, which works for detection and classification but not for language models with billions of parameters. The Hailo-10H adds memory on the module to hold those models. It uses the same Dataflow Compiler and HailoRT runtime as the Hailo-8 and Hailo-8L, so a project can move between them.
It ships as an M.2 module that fits the M key slot most industrial PCs already have for an NVMe SSD, which makes it one of the simplest ways to add generative AI to an existing x86 or Arm computer.
What it is designed to run
Hailo aims the 10H at generative workloads that need to stay on site, either because there is no reliable connection or because the data must not leave the building.
-
//
Generative AI. Runs small language and vision language models on site, such as models in the 1 to 3 billion parameter range.
-
//
Vision. Still runs detection, segmentation and pose models, so one module can handle both vision and language tasks.
-
//
Speech. Speech recognition and text-to-speech for voice interfaces on machines and kiosks.
-
//
On-module memory. 8 GB LPDDR4 holds the model, leaving system memory free for the application.
Hailo-10H vs Hailo-8 and Hailo-8L
All three use the same software, so the choice comes down to the model. Vision-only projects usually stay on the Hailo-8 family; the Hailo-10H is for models that need memory.
| Accelerator | AI performance | On-module memory | Form factor | Best for |
|---|---|---|---|---|
| Hailo-8L | Up to 13 TOPS (INT8) | None | M.2 module | Entry vision AI, one or two cameras |
| Hailo-8 | Up to 26 TOPS (INT8) | None | M.2 module or PCIe card | Multi-camera vision AI |
| Hailo-10H | Up to 40 TOPS (INT4) | 8 GB LPDDR4 | M.2 M key module | Generative AI plus vision |
The precisions differ, so the TOPS figures are not directly comparable: the 10H’s figure is quoted at INT4, a lower precision than the INT8 figures for the Hailo-8 family. Our Hailo-8 guide compares the three Hailo chips in more depth.
How it fits in a system
The Hailo-10H connects to the host processor over PCI Express through the M.2 slot, draws a few watts and is passively cooled by the system, so it suits fanless enclosures. The host CPU runs the application and passes requests to the module through HailoRT; the model weights stay in the module’s own memory. Our guide to M.2 slots and keys explains which slots can take an accelerator.
Each module takes one M key slot, often the one an NVMe drive would use, so the number of slots limits how many modules a system can carry. Some AMD Ryzen AI 300 industrial PCs in our range take up to two Hailo-10 modules beside the processor’s own NPU, and DIN rail systems on the Ryzen Embedded 8845HS do the same with an expansion board. As a general AI accelerator, it sits between an integrated NPU and a GPU.
Software and model support
Models are compiled for the chip with the Hailo Dataflow Compiler and run through the HailoRT runtime on Linux or Windows. Hailo publishes a model zoo of pre-compiled vision and generative models, which is the quickest way to test a use case before compiling your own. Language models are quantized to 4-bit weights to fit the module’s memory and use its INT4 performance, so accuracy should be checked on your own prompts and data before deployment. In a system with two modules, HailoRT can run a separate model on each, for example a language model on one and a vision model on the other.
Where the Hailo-10H is used
Most uses pair a language or vision-language model with cameras or microphones already on the machine, and keep every query on site.
-
//
Operator assistants. A panel PC or HMI answers spoken or typed questions about a machine from its manuals and logs, without an internet connection.
-
//
Scene description. A vision language model describes what a camera sees in plain text, for security, inspection or accessibility.
-
//
Smart search of video. Footage is tagged with descriptions as it is recorded, so staff can search it in plain language later.
-
//
Retail and kiosks. Voice and text interfaces that run on the device, so customer data stays in the building.
Quick answers on the Hailo-10H
Hailo-10H: common questions
How many TOPS is the Hailo-10H?
Up to 40 TOPS at INT4. The figure is quoted at 4-bit precision, which is lower than the INT8 figures usually given for the Hailo-8.
Is the Hailo-10H the same as the Hailo-10?
Effectively, yes. Hailo-10 is the family name and the Hailo-10H is the part found on M.2 modules, so product listings often use either name.
Can the Hailo-10H run an LLM?
Small ones, yes. Its 8 GB of memory holds quantized language models of a few billion parameters. Larger models need a GPU or a Jetson with more memory.
Does the Hailo-10H work with a Raspberry Pi?
Yes. Raspberry Pi’s AI HAT+ 2 uses the Hailo-10H. Industrial systems use the M.2 module in an x86 or Arm computer.
Does it replace a GPU?
For small models, often. It runs small language and vision models at a fraction of a GPU’s power. Larger models and many concurrent users still need a GPU.
Our Hailo AI accelerator computers show the systems we build around these modules. We fit the Hailo module, drivers and HailoRT in each system as standard and validate it before it ships.