8W
TYPICAL POWER
EdgeAgent E8 is a fanless edge AI appliance built on the Rockchip RK3588. Run 8 channels of 1080p video analytics and small language models entirely on-premises — no cloud bill, no fan, nothing to service.
Four things it does well. We'll tell you what it doesn't do in the next section.
Run 8× 1080p RTSP streams with real-time object detection on every stream — 20–25 FPS each, under 80 ms end-to-end. YOLOv8n detection at up to 110 FPS at 640×640 INT8.
Ships with the DeepSeek Harness (dsh) agent runtime preinstalled and configured. Schedule agent jobs, connect tools over MCP, automate document workflows — on a box that idles at 8 W instead of 45 W.
Sensitive documents and camera feeds are processed locally. Nothing is uploaded. When a task genuinely needs a large model, dsh escalates to a cloud API — and you control which tasks are allowed to.
A real HDMI capture input, not just output. Build video-wall controllers, kiosk capture and set-top-box applications without adding a camera.
We publish these because a returned product costs us more than a lost sale.
| Question | Honest answer |
|---|---|
| Can it run ChatGPT / DeepSeek / Llama-70B locally? | No. There is no discrete GPU. Local models are limited to 1.5B–4B quantized. Large-model reasoning routes to a cloud API you configure. |
| How fast is the local AI? | Qwen2.5-1.5B: 5–8 tok/s (usable). 3B: 2–3 tok/s (slow). 7B: 0.5–1 tok/s — don't buy it for this. |
| Is it a Jetson replacement? | No. If your code depends on CUDA or TensorRT, buy a Jetson. RK3588 uses the RKNN toolchain and porting is real work. |
| Can I train models on it? | No. Inference only. |
| Pose estimation or instance segmentation? | Not comfortably. Detection and classification are where it's strong. |
The 8 GB model runs video analytics and 1.5B–3B models comfortably. If you want the document assistant or local LLM work, choose the 16 GB model — the 8 GB will technically run them, but you'll be disappointed.
And who should buy something else.
| ✅ Good fit | ❌ Wrong fit |
|---|---|
| Security / CCTV integrators wanting on-prem analytics with no monthly fee | Anyone who needs to run 70B models locally |
| Factory automation with offline inference requirements | Teams with an existing CUDA / TensorRT codebase |
| Self-hosters wanting a silent, always-on AI box | Anyone wanting a plug-and-play consumer device with an app store |
| Kiosk / digital signage with 8K decode + HDMI capture | High-framerate pose estimation or segmentation |
| SMBs needing on-prem document AI for sensitive data | Model training |
Verified on production hardware.
| SoC | Rockchip RK3588, 8 nm |
| CPU | 8-core: 4× Cortex-A76 @ 2.4 GHz + 4× Cortex-A55 @ 1.8 GHz |
| NPU | 6 TOPS INT8 (triple-core, INT4 / INT8 / INT16 / FP16) |
| GPU | Mali-G610 MP4 |
| Video | 8K@60 H.265/VP9 decode, AV1 hardware decode, 2× encoder cores |
| RAM | 8 GB LPDDR4 · 16 GB recommended for local LLM work |
| Storage | 64 GB eMMC · 128 GB / +NVMe options available |
| Network | 2× Gigabit Ethernet · Wi-Fi + Bluetooth optional |
| USB | 2× USB 3.0 Type-C (one is USB-C PD power in), 1× USB 3.0 host, 4× USB 2.0 |
| Display | 1× HDMI out · 1× DisplayPort out · 1× HDMI input |
| Power | USB-C PD, ~8 W typical |
| Cooling | Fanless, passive |
| OS | Ubuntu 22.04.5 LTS (aarch64), Kernel 6.1 |
| Software | dsh agent runtime preinstalled + 3 working demos |
| Certification | CE (EU) · FCC (USA) · UKCA (UK) · SASO/SABER (Saudi) · ECAS (UAE) |
CAN FD, MIPI CSI camera inputs and SATA are disabled in the default firmware. If your project needs them, tell us before ordering — enabling them requires firmware work and we'll scope it with you.
For the included demos, no — they start on boot. For custom work, yes, you'll be on an Ubuntu terminal. We include a written quick start and answer questions on WhatsApp.
If they output RTSP or ONVIF, almost certainly. Send us a model number before ordering and we'll confirm.
The box arrives configured. Power it, connect Ethernet, and the demos are running. Getting your model on it is the real work — budget days, not minutes, for RKNN conversion.
Yes. We quote FOB or DDP depending on your region. We'll give you the exact HS code and the paperwork you need.
Because it's a different tool. RK3588 has a weaker ecosystem and a less mature toolchain. You trade some performance and a lot of ecosystem maturity for roughly half the price and a third of the power draw.
Yes — anything convertible to RKNN. We'll help you scope the conversion, which is usually the hardest part of an edge project.
Send us what you're trying to build. If this box is wrong for it, we'll say so — and point you at what you actually need.