AI marketplace for Raspberry Pi fleets
AI Marketplace Control Plane
Ship Hugging Face models as managed container recipes, monitor edge inference quality, and remotely support Raspberry Pi AI HAT 2 deployments.
Fleet Health
Retail Vision Europe
Contoso Retail · fleet-retail-eu
production
vision
Berlin Store 01
Berlin · Raspberry Pi AI HAT 2
online
17.8 FPS · 58.0 °C
CPU 44.0% · Mem 52.0%
Paris Store 08
Paris · Raspberry Pi AI HAT 2
degraded
10.2 FPS · 74.0 °C
CPU 79.0% · Mem 88.0%
Simulated Pi 01
Docker Lab · Raspberry Pi AI HAT 2
online
17.2 FPS · 64.8 °C
CPU 60.3% · Mem 45.3%
asdasdas
asdasd · Raspberry Pi AI HAT 2
online
19.5 FPS · 63.5 °C
CPU 0.3% · Mem 6.3%Marketplace Catalog
Agl Computervision Experiments
unknown · Alejandro-sin/agl-computervision-experiments
license:mit
region:us
- Container:
ghcr.io/example/alejandro-sin-agl-computervision-experiments:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Agnuxo Qwen2_0.5B_Spanish_English_Raspberry_Pi5_16Bit Gguf
text-generation · featherless-ai-quants/Agnuxo-Qwen2_0.5B_Spanish_English_raspberry_pi5_16bit-GGUF
gguf
text-generation
base_model:Agnuxo/Qwen2_0.5B_Spanish_English_raspberry_pi5_16bit
base_model:quantized:Agnuxo/Qwen2_0.5B_Spanish_English_raspberry_pi5_16bit
endpoints_compatible
region:us
- Container:
ghcr.io/example/featherless-ai-quants-agnuxo-qwen2-0-5b-spanish-english-raspberry-pi5-16bit-gguf:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Agnuxo_ _Qwen2_0.5B_Spanish_English_Raspberry_Pi5_16Bit Awq
unknown · RichardErkhov/Agnuxo_-_Qwen2_0.5B_Spanish_English_raspberry_pi5_16bit-awq
safetensors
qwen2
4-bit
awq
region:us
- Container:
ghcr.io/example/richarderkhov-agnuxo-qwen2-0-5b-spanish-english-raspberry-pi5-16bit-awq:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Anzhcs_Yolos
object-detection · Anzhc/Anzhcs_YOLOs
ultralytics
pytorch
YOLOv8
art
Ultralytics
object-detection
- Container:
ghcr.io/example/anzhc-anzhcs-yolos:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Computer Vision Face Detection
unknown · Chetan007/computer-vision-face-detection
region:us
- Container:
ghcr.io/example/chetan007-computer-vision-face-detection:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Computer_Vision_Class
unknown · andresq99/Computer_Vision_Class
license:apache-2.0
region:us
- Container:
ghcr.io/example/andresq99-computer-vision-class:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Computervision
unknown · NehaBardeDUKE/computervision
license:bigscience-openrail-m
region:us
- Container:
ghcr.io/example/nehabardeduke-computervision:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Llama 3.2 11B Vision Instruct
image-text-to-text · meta-llama/Llama-3.2-11B-Vision-Instruct
transformers
safetensors
mllama
image-text-to-text
facebook
meta
- Container:
ghcr.io/example/meta-llama-llama-3-2-11b-vision-instruct:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Nomic Embed Vision V1.5
image-feature-extraction · nomic-ai/nomic-embed-vision-v1.5
transformers
onnx
safetensors
nomic_bert
feature-extraction
image-feature-extraction
- Container:
ghcr.io/example/nomic-ai-nomic-embed-vision-v1-5:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Phi 4 Reasoning Vision 15B
image-text-to-text · microsoft/Phi-4-reasoning-vision-15B
safetensors
phi4-siglip
multimodal
vision-language
reasoning
math
- Container:
ghcr.io/example/microsoft-phi-4-reasoning-vision-15b:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Phi 4 Reasoning Vision 15B Gguf
text-generation · jamesburton/Phi-4-reasoning-vision-15B-GGUF
gguf
phi4
phi-4
quantized
llama-cpp
ollama
- Container:
ghcr.io/example/jamesburton-phi-4-reasoning-vision-15b-gguf:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Qwen2_0.5B_Spanish_English_Raspberry_Pi5_16Bit
text-generation · Agnuxo/Qwen2_0.5B_Spanish_English_raspberry_pi5_16bit
transformers
safetensors
qwen2
text-generation
text-generation-inference
unsloth
- Container:
ghcr.io/example/agnuxo-qwen2-0-5b-spanish-english-raspberry-pi5-16bit:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Realistic_Vision_V6.0_B1_Novae
text-to-image · SG161222/Realistic_Vision_V6.0_B1_noVAE
diffusers
license:creativeml-openrail-m
endpoints_compatible
diffusers:StableDiffusionPipeline
region:us
- Container:
ghcr.io/example/sg161222-realistic-vision-v6-0-b1-novae:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Stockmarket Pattern Detection Yolov8
object-detection · foduucom/stockmarket-pattern-detection-yolov8
ultralytics
tensorboard
v8
ultralyticsplus
yolov8
yolo
- Container:
ghcr.io/example/foduucom-stockmarket-pattern-detection-yolov8:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Tiny Llama_Spanish_English_Raspberry_Pi5_16Bit
text-generation · Agnuxo/tiny-llama_Spanish_English_raspberry_pi5_16bit
adapter-transformers
safetensors
llama
text-generation
text-generation-inference
transformers
- Container:
ghcr.io/example/agnuxo-tiny-llama-spanish-english-raspberry-pi5-16bit:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Traffic_Management_Using_Computer_Vision
unknown · ar1725/Traffic_Management_Using_Computer_Vision
region:us
- Container:
ghcr.io/example/ar1725-traffic-management-using-computer-vision:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Valorant Yolo11M
object-detection · jparedesDS/valorant-yolo11m
tensorboard
yolo11
valorant
object
detection
object-detection
- Container:
ghcr.io/example/jparedesds-valorant-yolo11m:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Whisper Tiny Edge Speech
speech-to-text · openai/whisper-tiny
audio
edge
fleet-ready
- Container:
ghcr.io/example/whisper-tiny-rpi:latest - Recipe: recipes/pi-speech-inference.recipe.yaml
- RAM: 6.0 GB · Storage: 12 GB
- Bench: 1.4 FPS / 711.0 ms / 6.1 W
YOLOv8 Nano Edge Vision
object-detection · ultralytics/yolov8n
vision
raspberry-pi
real-time
- Container:
ghcr.io/example/yolov8n-rpi:latest - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 19.2 FPS / 52.0 ms / 7.8 W
Yolo11
unknown · Ultralytics/YOLO11
ultralytics
en
zh
ja
ru
de
- Container:
ghcr.io/example/ultralytics-yolo11:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Yolov11 License Plate Detection
object-detection · morsetechlab/yolov11-license-plate-detection
ultralytics
onnx
computer-vision
object-detection
license-plate
yolov11
- Container:
ghcr.io/example/morsetechlab-yolov11-license-plate-detection:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Yolov8
object-detection · Ultralytics/YOLOv8
ultralytics
tracking
instance-segmentation
image-classification
pose-estimation
obb
- Container:
ghcr.io/example/ultralytics-yolov8:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Yolov8M Scene Classification
image-classification · keremberke/yolov8m-scene-classification
ultralytics
tensorboard
v8
ultralyticsplus
yolov8
yolo
- Container:
ghcr.io/example/keremberke-yolov8m-scene-classification:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Yolov8M Visdrone
object-detection · mshamrai/yolov8m-visdrone
ultralytics
tensorboard
v8
ultralyticsplus
yolov8
yolo
- Container:
ghcr.io/example/mshamrai-yolov8m-visdrone:hf-sync - Recipe: recipes/pi-vision-inference.recipe.yaml
- RAM: 4.0 GB · Storage: 16 GB
- Bench: 8.0 FPS / 125.0 ms / 7.0 W
Rollout History
Create Rollout
Device Metrics
Fleet Overview — latest readings
Enroll a New Device
What happens
1. Fill in the device details and click Generate Enrollment Command.
2. A one-time token is issued (no password stored).
3. Run the generated command on your device — it enrolls, then opens a persistent outbound connection to this control plane.
4. The device appears in Fleet Health within seconds.
Docker test device
docker compose -f docker-compose.rpi.yml up --build -d ssh pi@localhost -p 2223 # password: raspberry
Real Raspberry Pi
pip install httpx websockets asyncssh curl -O http://<control-plane>/static/agent.py
Request Tunnel Access
How tunnels work
A signed tunnel request token is issued by the control plane and validated by the edge agent. The token embeds device ID, requestor, purpose, and expiry.
Once you have a signed request token, send it to the agent:
curl -X POST http://localhost:8100/tunnel/open \
-H "Content-Type: application/json" \
-d '{
"requested_by": "you",
"purpose": "...",
"duration_minutes": 30,
"signed_request": "<token>"
}'
Administration
| ID | Name | Actions |
|---|
| Name | Tenant | Role | Last Login | Actions |
|---|
Sync from Hugging Face
Sync Notes
Sync fetches models from the Hugging Face API and upserts them into the marketplace catalog. Existing models are updated; new ones are added with default Pi benchmark estimates.
Set AIM_HUGGINGFACE_TOKEN in .env to authenticate for higher rate limits and private models.
Imported models use "source": "huggingface" to distinguish them from seed models.