Connect to AIHub

This chapter introduces the two sub-tabs of the Workbench Model Tab (Model Market + User Models), and how to pick a model in preparation for deploying it to the board.

Prerequisites

Item	Status
Workbench logged in to AIHub	✅ See ch01
Pico-G1 connected	✅ See ch02, ch03
AIHub training platform account	Registered (aihub.shimetapi.cn)

Step 1: Open the Model Tab

Click the 5th tab at the top, "Model":

The "Model Market" sub-tab of the Model Tab — a grid of community and official model cards

You'll see two sub-tabs:

Sub-Tab	Data Source	Visible To	Who Can Download
Model Market	Company official + community user publications	All logged-in users	All logged-in users
User Models	Models you trained yourself on AIHub	Current user only	Yourself (can also publish to the community)

Step 2: Browse the Model Market

Top Filters

Filter	Options
Format	All / pt / onnx / xmm / rknn

Model Card Fields

Each model card shows:

Field	Description
Model name	Short identifier (e.g. `mask` / `yolov8n`)
Author (community models)	Publisher's username
Type badge (official models)	Official / Community
Format	PT / ONNX / RKNN / XMM
Size	Actual file size for each format
Input size	Model inference input (e.g. 640×640)
Publish date	e.g. 2026/8/24
mAP50 (some community models)	Mean average precision
Deployed/Downloads	Deployment status and download count

Community Model Examples

Model	Author	mAP50	Deployed
mask (based on yolov8n, 100 epochs)	GOBLIN	95.2%	Deployed
ball_new	GOBLIN	—	—
Human Five-Category Action Detection	claude code	98.7%	—
number_recognition (digits 0-9)	GOBLIN	80.9%	—

Official Model Examples

Model	Base Model
yolov8n	yolov8n.pt
yolov5n	yolov5n.pt
naruto_hand_seal_detection	naruto_hand_seal_detection.pt

Step 3: Browse User Models

Switch to the User Models sub-tab:

The User Models tab — a list of model cards you trained yourself

Field	Description
Model name (e.g. `2026.8.24`, `plus`, `gaijinban`)	Your training versions
Format	PT / ONNX / RKNN / XMM
Input size	640×640
Publish date	Training date

Every model card has a "Deploy to Board" button (see ch05).

Step 4: Choosing a Model (First-Time Recommendation)

For first use, we recommend a lightweight demo from the official models:

Recommendation	Model	Best For
⭐ First	yolov8n	General object detection, 80 classes (COCO)
Second	yolov5n	Faster, slightly lower accuracy

Step 5: Model Training (Optional)

If you already have a model, go straight to ch05 to deploy it.

To train your own model:

Click the "Model Training" link in the upper-right corner
The browser opens the training page at aihub.shimetapi.cn
Finish training and export the model

After export, the model automatically appears in the Workbench User Models sub-tab.

See AIHub Docs → Training Guide for details.

Step 6: Model Formats

Format	Applicable Board	Purpose
pt	—	Raw PyTorch weights, for continued training
onnx	—	Generic intermediate format, convertible
xmm	Pico-G1 (GK7206)	Model package runnable on the board
rknn	RK3588 (coming)	Rockchip NPU acceleration

Pico-G1 currently only supports deploying the .xmm format.

Step 7: Model Deployment Permissions

Deploying a model requires no special permissions. However:

Community models belong to their publishers and follow the licenses set by the publisher (see the model detail page)
Official models follow the company's commercial license

FAQ

Symptom	What to Check
Model Tab loads slowly	Check the PC network; the first load fetches the model list
Can't find a model you trained	The AIHub training task wasn't exported, or the format isn't .xmm
Model card shows only PT/ONNX, no XMM	XMM format wasn't selected when exporting after training

ch05 First Inference: deploy the chosen model to the board + view the feed in the Vision Tab
User Guide → Model Market: advanced filters, batch operations, model details
AIHub Docs → Training & Export: train your own model