Hailo-10H

schematics

none

toolchain

partial (open runtime + MIT GenAI model zoo)

datasheets

public (June 2025)

blobs

required (proprietary silicon/firmware)

retail

open_market (M.2 modules purchasable)

Proprietary silicon with published datasheets and an open generative-AI model zoo, but the silicon and the full toolchain are closed. The GenAI model zoo is MIT-licensed, as GitHub's own repository metadata records. HailoRT is MIT and LGPL-2.1, while the developer-zone documentation sits behind registration and the Dataflow Compiler behind a customer relationship, which is the closed half of the toolchain. Firmware blobs are required because HailoRT's driver ships the device firmware. The published documentation is AAEON's UP page, which carries the full specification table for the module, and retail availability is that same page plus Geniatech shipping an independent Hailo-10 module. hailo.ai itself refuses every automated request, so all of this rests on substitute sources.

• https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-ai-acceleration-module/ recorded 2026-06-22

  official module product page for Hailo-10H. Not re-fetchable - hailo.ai answers HTTP 403 to every automated request under every header set tried - so it keeps its recorded read

• https://api.github.com/repos/hailo-ai/hailo_model_zoo_genai recorded 2026-08-14

  "Model zoo for Gen AI models for Hailo products" with license spdx_id MIT, 100 stars

• https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md recorded 2026-08-14

  "The master branch supports only the Hailo-10 and Hailo-15 device families"; MIT for libhailort/pyhailort/hailortcli and LGPL 2.1 for hailonet; developer-zone documentation "registration is required"; the PCIe driver ships device firmware

• https://up-board.org/hailo-10h-m2-ai-acceleration-module/ recorded 2026-08-14

  AAEON/UP's published spec table for the module - chipset, AI performance, interface, memory and speed, power, supported OS, host architectures, frameworks and model types. What it does NOT publish is any board file - "Package Contents - 1 x Quick Start Guide" is the whole of what ships, and no schematic, gerber or BOM is offered anywhere on the page, which is the evidence for schematics:none.

• https://www.geniatech.com/product/aim-m-h10/ recorded 2026-08-14

  a second vendor's purchasable Hailo-10 M.2 module - "Up to 40 TOPS AI performance", "M.2 Key-M form factor"

Newer generative-AI part; purchasable M.2 modules and a dedicated GenAI repo, but smaller install base than the Hailo-8. The part is productized and purchasable through at least two independent module vendors (AAEON/UP and Geniatech), and it is visibly the smaller surface next to the Hailo-8 - the GenAI model zoo carries 100 stars against HailoRT's 215 - which is why niche at level 3, one below the Hailo-8, is where it sits.

• https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-ai-acceleration-module/ recorded 2026-06-22

  productized, purchasable M.2 generative-AI module. Not re-fetchable - hailo.ai answers HTTP 403 to every automated request - so it keeps its recorded read

• https://up-board.org/hailo-10h-m2-ai-acceleration-module/ recorded 2026-08-14

  AAEON/UP shipping the Hailo-10H M.2 module as a listed product with full specifications

• https://www.geniatech.com/product/aim-m-h10/ recorded 2026-08-14

  "Hailo-10 M.2 AI Accelerator Card" - a second independent vendor productizing the part

• https://api.github.com/repos/hailo-ai/hailo_model_zoo_genai recorded 2026-08-14

  the dedicated GenAI repo at 100 stars, 20 forks, last pushed 2026-04-15 - real but well behind HailoRT's 215 stars and 2026-08-11 push

Edge-first part that runs generative models locally via a DDR interface, a step beyond CNN-only edge accelerators. AAEON/UP reproduces the recorded value exactly - "AI Performance 40 TOPS @ INT4 | 20 TOPS @ INT8" and "Hailo-10H includes a direct DDR interface, allowing it to scale for large models such as LLMs, VLMs, Stable Diffusion" - and Geniatech independently states up to 40 TOPS running LLMs and VLMs locally. That is one band above the Hailo-8, which is CNN/vision-only and bounded by on-die memory. hailo.ai itself is unfetchable, so both rest on substitutes.

• https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-ai-acceleration-module/ recorded 2026-06-22

  40 TOPS INT4 / 20 TOPS INT8, DDR interface, LLM/VLM support. Not re-fetchable - hailo.ai answers HTTP 403 to every automated request - so it keeps its recorded read

• https://up-board.org/hailo-10h-m2-ai-acceleration-module/ recorded 2026-08-14

  "AI Performance 40 TOPS @ INT4 | 20 TOPS @ INT8"; "Hailo-10H includes a direct DDR interface, allowing it to scale for large models such as LLMs, VLMs, Stable Diffusion"; "Supported AI Models Vision Models (> 150+), GenAI Models (LLM, VLM, Whisper)"

• https://www.geniatech.com/product/aim-m-h10/ recorded 2026-08-14

  "Up to 40 TOPS AI performance"; "Generative AI ready - Runs LLMs, VLMs, and Stable Diffusion locally"; "AI Model Types CNN, Transformer, Vision Transformer, Generative AI Models"