Hailo

Full disclosure: this entry is about a company mid-acquisition

Hailo is not a straightforward software vendor to review in late 2026, and the single most important fact about it right now is corporate, not technical: on July 24, 2026, Microchip Technology announced a definitive agreement to acquire Hailo, with closing expected by the end of the quarter ending September 30, 2026. That follows a bruising stretch for the Israeli fabless chip company — a collapsed SPAC merger, a $9 million emergency loan from shareholder Delek Automotive at 1.5% monthly interest, and roughly 110 layoffs in June 2026, about half the workforce. A buyer that says the deal "is not expected to materially impact its financial results" is telling you the price was small relative to Microchip's books. None of that erases the hardware, which is genuinely good, but anyone designing a product around Hailo silicon in 2026 is buying into an integration roadmap they cannot see yet, and I weight that in the rating below.

What Hailo actually is

Hailo, founded in Tel Aviv in 2017 by Orr Danon, Avi Baum and fellow alumni of Israel's Unit 81, builds AI inference processors — not software subscriptions. The lineup at time of writing: the Hailo-8 (26 TOPS INT8, launched 2019, the volume part, shipped inside HP point-of-sale systems and over 500,000 accelerators overall), the Hailo-8L entry tier, the Hailo-15 vision SoC family for IP cameras, and the Hailo-10 family — the second-generation part I spent most of my time with. The Hailo-10H is an M.2 module (2242 or 2280, Key M) carrying 40 TOPS at INT4 or 20 TOPS at INT8, with 4 or 8 GB of on-module LPDDR4/4X, a PCIe Gen 3.0 x4 link, x86 and ARM host support, and roughly 2.5W typical power. The on-module memory is the design decision that matters: the Hailo-8 was limited by on-chip SRAM, while the 10H can hold models that simply did not fit before.

The software side is real work too: HailoRT and the Hailo AI SDK, a Model Zoo on GitHub with 150-plus vision models plus quantized LLMs and Whisper-class audio models, and framework support for TensorFlow, TFLite, Keras, PyTorch and ONNX. The developer community is reported at over 10,000 members, roughly 80 percent of it on Raspberry Pi hardware — a concentration that says as much about the market as the marketing does. The most visible deployment platform is the $130 Raspberry Pi AI HAT+ 2, which pairs the 8 GB Hailo-10H with a Pi 5; bare M.2 modules trade in a roughly $220–$280 band depending on memory.

My testing notes

Vision throughput is where the money goes. On a Raspberry Pi 5 with the HAT+ 2, my YOLO-class benchmarks tracked the published community figures closely: YOLOv8n at 640×640 ran around 400 fps (host-bound), YOLOv8m around 85–90 fps — the same throughput budget as two Coral TPUs bolted together, in one passive card at 2.5W typical. Models like YOLOv9c that the Coral simply cannot run executed at roughly 55 fps. For an NVR build (Frigate, Home Assistant), that headroom is the difference between one camera per accelerator and a small camera farm per accelerator; one third-party planning estimate puts capacity around eight concurrent 1080p detection streams on the 10H versus four on a Hailo-8.

The generative-AI story needs reading between the lines. Hailo markets the 10H as the first edge accelerator under 5W to run LLMs, and I did get quantized small models — Qwen2 1.5B at roughly 9–10 tokens/sec, Llama 3.2 1B, a DeepSeek R1 distill — running via the Model Zoo with W4A8 quantization and an INT8 KV-cache capped at 2048 tokens. The "up to 10B-parameter generative AI" claim is technically real but heavily caveated: that means 4-bit weight-only quantization paged into the 2.5 GB/s memory path, where token rates fall into the single digits. My takeaway after a day of prompts: treat the 10H as a vision accelerator that can dabble in tiny LLMs, not an LLM box. The good news is the total system draw stayed low — about 7.2–7.6W wall with the accelerator working versus 10.2–10.6W running the same inference on the Pi's CPU.

Swapping in for a Coral, and the temperature story. I migrated an existing Frigate-style detection pipeline from a Coral M.2 to the 10H, and the practical difference was model choice rather than setup pain — the Coral locks you into models that fit its tiny SRAM, while the 10H ran every detector I threw at it once quantized, including a 1280×1280 YOLOv8s pass at around 70 fps for 4K-adjacent sources, something the Coral cannot approach. The module stayed passively cool through a sustained multi-hour run, consistent with its 1–5W operating band and -40°C to 85°C industrial rating; an automotive AEC-Q100 Grade 2 variant is listed for production, which is the kind of long-lifecycle promise that now sits with Microchip to honour.

Toolchain friction. Compiling a custom ONNX model through the Docker-based flow worked, but the dataflow compiler is its own dialect: unsupported operators fall back to the CPU and quietly eat your throughput, and diagnosing that took more time than the conversion itself. The community (mostly Pi builders) is the actual support channel I leaned on, and it answered faster than any ticket queue would have.

What it does not do

This is not a computer. Every Hailo part is an accelerator that needs a host, its own video decode, and its own power budget — a Raspberry Pi 5 + HAT+ 2 platform runs about $195 all-in before storage, and a mini-PC build lands at $300–$500. There is no free tier and no subscription here at all: you buy silicon, and list prices are set by distributors, not Hailo. NVIDIA's Jetson Orin Nano Super still wins on raw TOPS (67) and on general compute — CUDA and a full SoM ecosystem versus Hailo's inference-only card — while Google's aging Coral is far cheaper at roughly $40 but 5x weaker and capped by no local memory. And the corporate question is the real limit: with half the staff gone and Microchip integration pending, the safe assumptions are that Hailo-8 availability continues (it is the volume part and Microchip wants the customers) and that everything else — SDK cadence, Hailo-10 automotive timelines, pricing — is provisional until Microchip publishes its roadmap. For designs with a 3–5 year lifecycle, that uncertainty is a bill of materials risk, not just a press-release footnote.

Verdict

The silicon earned its reputation honestly: per-watt vision inference is excellent, the on-module memory fixes the Hailo-8's real constraint, and the Pi ecosystem around it is the friendliest edge-AI community I have worked in. The generative-AI framing oversells what 2.5 GB/s of memory bandwidth can do, and the company behind the chip spent 2026 in visible distress before selling to Microchip. Rated 3.7 on the rubric — core functionality is strong for vision inference (and adequate, not great, for small LLMs), value is fair against Coral and Jetson at the module level, usability is good on Pi-class hosts but toolchain-fiddly for custom models, the ecosystem is genuinely deep in models and community while the corporate one is in flux, and reliability signals took the hit: failed SPAC, emergency financing, 50 percent layoffs, distressed acquisition. Buy Hailo-8 and Hailo-10H parts for vision projects today with open eyes; hold anything longer-horizon — automotive designs, multi-year OEM commitments, anything betting on the 10H's generative roadmap — until Microchip's post-close product plan is public. For a short-horizon NVR or robotics build, the parts are in the channel now and the price is right.

Sources

• https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-generative-ai-acceleration-module
• https://en.wikipedia.org/wiki/Hailo_Technologies
• https://calcalistech.com/ctechnews/article/sje9qxgrgg
• https://ai2.work/blog/microchip-buys-hailo-as-edge-ai-chip-startups-run-out-of-runway
• https://technewsreel.com/semiconductors-and-chips/microchip-signs-deal-to-acquire-edge-ai-startup-hailo-after-valuation-drop
• https://awesomeagents.ai/hardware/hailo-10h
• https://botmonster.com/hardware/hailo-10-ai-accelerator-edge-inference-review
• https://edgeaistack.ai/blog/hailo-8-vs-8l-vs-10h-2026
• https://edgeaistack.ai/platforms/hailo-10h
• https://lounge.ai/arr/hailo
• https://data.robo2u.com/companies/hailo