Can the RK3588 run chat and vision chat?

Not at present. The reComputer RK3588 Series gets the object detection preset only; LLM chat and vision chat support the reComputer RK3576 Series.

Does it work offline, and does data leave the device?

After deployment it runs offline; the first deploy needs internet access to pull a 1-4GB image. All inference runs on the box's NPU, and images and conversations are not sent to any cloud.

It depends on the model size you pick. Nano is the fastest detector and Medium the most accurate; 1.5B chat answers fastest and 7B reasons better. Vision chat takes 1-2 minutes to load before it is ready.

How do I connect it to my own application?

Detection is a REST API that returns class, confidence and box coordinates for an uploaded image. Chat and vision chat are OpenAI-compatible, so an existing OpenAI SDK only needs the device IP as its base URL. The vision API ships interactive docs you can try in a browser.

Can I swap in my own model or custom classes?

The presets switch between the listed model sizes only. The detector recognizes 80 common classes; custom classes mean training your own model and converting it to the Rockchip NPU format.

Can all three presets run on one device?

Yes. They are independent, use ports 8000, 8001 and 8002, and can be deployed together on one RK3576 box.

What should we check on model licensing for a commercial product?

Each set of weights keeps its own licence. Ultralytics distributes YOLO 11 weights under AGPL-3.0, and DeepSeek-R1 and Qwen2.5-VL carry their own terms. Before shipping a closed product, confirm the licence of the model you use or swap in a compatibly-licensed one.

AI Lab — Edge AI Model Hub on Rockchip

One place to try popular AI models on reComputer RK3576/RK3588 — object detection, chat, and image understanding, all running locally.

SenseCraftBeginner15minOn-Device LMAI Agent Ready

Reusable Site Templates

Teams that want to try popular models on their own Rockchip box before deciding what to build.

Object detection for existing cameras or an NVR — the NVR posts a frame and gets boxes for 80 common object classes back.
A chat assistant on a closed or offline network — point an existing OpenAI client at the device IP; conversations stay inside.
First-pass image review — upload a screenshot or photo and ask about it in plain language.

Not a fit: projects that need a finished app UI, since this ships APIs. Chat and vision chat run only on the reComputer RK3576 Series.

How it works

One Rockchip box is enough; the models run on its built-in NPU. Add a USB camera only for live video detection. Your application calls the APIs on the box over the LAN.

Only live video detection needs a camera; one USB camera plugged into the box is enough.

Works without a camera: the image upload API is always available.
Chat and vision chat need no camera: images go in through the API.

The three presets are independent and can run together on one RK3576 box, on ports 8000, 8001 and 8002.

Preset	Model options	reComputer RK3576 Series	reComputer RK3588 Series	Memory
Object detection	YOLO 11 Nano / Small / Medium	Yes	Yes	—
LLM chat	DeepSeek-R1 1.5B / 7B, 5 quantizations	Yes	—	4GB for 1.5B, 8GB+ for 7B
Vision chat	Qwen2.5-VL 3B	Yes	—	8GB+

Each preset takes 3-10GB of disk, depending on the model chosen.

How well it works

The table lists what the three presets provide on the reComputer RK3576 Series, and their limits.

Preset	What it does	Limit
Object detection	Finds 80 common object classes in images or live video	Only these 80 classes; new ones need a different model
LLM chat	Local text chat, OpenAI-compatible	7B needs 8GB+ RAM; 1.5B answers faster
Vision chat	Describes images and answers questions about them	Needs 8GB+ RAM; first load takes 1-2 minutes

Speed follows the model size: Nano is the fastest detector, Medium the most accurate. First start downloads a 1-4GB image and loads the model.

What you get out of it

Once deployed, these are the interfaces.

Interface	What comes out	Notes
HTTP 8000	Image detection: class, confidence and box coordinates	Confidence threshold is adjustable
HTTP 8000	MJPEG video feed with detection boxes drawn	Needs a USB camera
HTTP 8001	OpenAI-compatible chat completions	Streaming supported
HTTP 8002	OpenAI-compatible vision API	Text plus image input, interactive API docs

Every request stays on the LAN; images and conversations are not sent to any cloud.

How to deploy

Check the setup and its product list, then follow the deployment guide.

Choosing

Object detection only, with a stronger NPU → reComputer RK3588 Series, running the detection preset.
Detection, chat and vision chat together → reComputer RK3576 Series; all three presets can share the box.
7B chat or vision chat → an RK3576 model with 8GB of RAM or more.

Before deploying: the box accepts SSH and has Docker installed, and the first deploy needs internet access to pull images.

Hardware Sizing & Deployment Setup

Select site conditions to generate recommended bundles and guides, or let an AI assistant estimate hardware and deploy for you.

Self-service Consult & Deploy

Deployment Options

Object Detection (YOLO 11)

Runs YOLO 11 on RK3576 or RK3588, detecting 80 objects; choose model size at deploy time.

2 steps · ~15min

Deployment Guide →

Edge Device

recomputer_rk3576

recomputer_rk3588

FAQ

One reComputer RK3576 Series or RK3588 Series box. 1.5B chat runs with 4GB of RAM; 7B chat and vision chat need 8GB or more. Reserve 3-10GB of disk per preset.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !

AI Lab — Edge AI Model Hub on Rockchip

Reusable Site Templates

How it works

How well it works

What you get out of it

How to deploy

Choosing

Hardware Sizing & Deployment Setup

Object Detection (YOLO 11)

FAQ

What hardware do I need, and what is the minimum?

Can the RK3588 run chat and vision chat?

Does it work offline, and does data leave the device?

How fast is it?

How do I connect it to my own application?

Can I swap in my own model or custom classes?

Can all three presets run on one device?

What should we check on model licensing for a commercial product?

AI Lab — Edge AI Model Hub on Rockchip

Reusable Site Templates

How it works

Camera (Optional)

Local Host

Applications

How well it works

What you get out of it

How to deploy

Choosing

Hardware Sizing & Deployment Setup

Object Detection (YOLO 11)

FAQ

What hardware do I need, and what is the minimum?

Can the RK3588 run chat and vision chat?

Does it work offline, and does data leave the device?

How fast is it?

How do I connect it to my own application?

Can I swap in my own model or custom classes?

Can all three presets run on one device?

What should we check on model licensing for a commercial product?