Vision input
Arducam 16MP Autofocus Camera Module
IMX519 16MP autofocus camera with ABS case for Raspberry Pi models.
View on AmazonEquipment guide · sensor body + brain
Build the body. Then give it a brain.
JARVIS is intentionally split in two: a Raspberry Pi 5 sensor body for presence, and a local GPU brain for cognition, memory, voice, learning, and governance.
Purchase list
Buy the exact body and brain equipment.
These are outbound Amazon links for the hardware stack. The brain machine should be treated as an offline AI processor, not a Windows desktop.
For the Senses
Raspberry Pi sensor body
The physical JARVIS presence: camera, touch display, mic, speaker, Pi 5, and Hailo acceleration.
Sensor computer
CanaKit Raspberry Pi 5 16GB Starter Kit PRO
16GB Pi 5 kit with 128GB storage edition for the always-on senses node.
View on AmazonPresence display
FREENOVE 5 Inch Touchscreen Monitor
800x480 IPS capacitive touchscreen over MIPI DSI for the Pi display surface.
View on AmazonVoice output
Portable Mini Sound Bar
Compact stereo speaker with enhanced bass for local TTS playback.
View on AmazonAudio input
USB Gooseneck Microphone
360 degree adjustable USB mic with mute button, LED indicator, and noise-canceling tech.
View on AmazonVision acceleration
Raspberry Pi AI HAT+2
Hailo-10H accelerator with 8GB on-board RAM and 40 TOPS class AI capability for Pi 5.
View on AmazonFor the Brain
Offline AI workstation parts
The brain machine runs cognition, memory, models, governance, and self-improvement locally.
CPU backbone
AMD Ryzen 9 desktop CPU
High-thread-count Ryzen 9 class processor for the brain and CPU-resident model lanes.
View on AmazonPremium GPU brain
NVIDIA GeForce RTX 4080 16GB
Strong baseline GPU for local STT, TTS, LLM residency, vision support, and fast iteration.
View on AmazonMore power
ASUS ROG Astral GeForce RTX 5090 BTF OC
32GB GDDR7 option for heavier local model residency and a bigger extreme-tier brain.
View on AmazonComplete brain option
Prebuilt workstation path
A ready-to-go tower option if you want the brain hardware assembled first, then reinstalled with Linux.
Prebuilt brain machine
CLX Horus Gaming PC
Ryzen 9 9950X3D, RTX 5080, 96GB DDR5, 4TB NVMe class build. Install Pop!_OS over Windows before using it as the brain.
View on AmazonBrain OS requirement
Install Pop!_OS over the PC used for the brain. Do not run the brain as a Windows install. Treat that machine as the offline local AI processor for JARVIS.
Sensor body
Pi 5, camera, mic, speaker, display, and Hailo acceleration.
The Pi is the senses. It does not own memory, personality, policy, self-improvement, or the LLM. It captures the room and streams events to the brain.
Thin sensor node
Raspberry Pi 5
Runs the senses layer: camera capture, Hailo vision inference, audio capture/playback, WebSocket transport, and the particle display.
Vision accelerator
Hailo-10H AI HAT+
Runs YOLOv8s person detection, SCRFD face detection, and YOLOv8s-Pose locally on the Pi so the brain receives structured perception.
Vision input
Pi camera
Feeds Picamera2 for person detection, pose, facial expression, face crops, and scene summaries.
Audio input
USB microphone
Captures 44.1kHz audio, resamples to 16kHz int16, and streams raw PCM to the brain over local WebSocket.
Voice output
Speaker
Plays back brain-synthesized TTS audio. The Pi does not run the language or speech intelligence.
Presence surface
7 inch display
Runs the JARVIS particle visualizer in kiosk mode and reflects bounded system state, not private dashboard internals.
Brain equipment
The GPU tier decides model residency, latency, and how much can stay awake.
The brain auto-detects NVIDIA VRAM, CPU threads, and RAM at startup. It then chooses LLM size, STT model, TTS device, vision availability, model keep-alive, and whether ancillary ML should live on CPU or GPU.
Hardware tiers
GPU VRAM selects the brain profile.
The brain auto-detects GPU VRAM at startup and selects model sizes, compute types, and memory strategy from seven tiers. Local-first guarantee: all core capabilities run entirely on local hardware.
| Tier | VRAM | LLM | Fast | Vision | STT | TTS | Keep-alive |
|---|---|---|---|---|---|---|---|
| minimal | <4 GB | qwen3:1.7b | qwen3:1.7b | disabled | tiny/int8 | none | 5m |
| low | 4-6 GB | qwen3:4b | qwen3:1.7b | disabled | small/int8 | none | 5m |
| medium | 6-8 GB | qwen3:8b | qwen3:4b | disabled | medium/int8_fp16 | kokoro_cpu | 5m |
| high | 8-12 GB | qwen3:8b | qwen3:4b | qwen2.5vl:7b | large-v3-turbo | kokoro_cpu | 10m |
| premium | 12-16.5 GB | qwen3:8b | qwen3:8b | qwen2.5vl:7b | large-v3/int8_fp16 | kokoro_gpu | 30m |
| ultra | 16.5-24.5 GB | qwen3:14b | qwen3:8b | qwen2.5vl:7b | large-v3/float16 | kokoro_gpu | always |
| extreme | 24.5 GB+ | qwen3:32b | qwen3:14b | qwen2.5vl:7b | large-v3/float16 | kokoro_gpu | always |
Ultra+ tiers pin models in VRAM permanently, eliminating cold-start latency. Premium uses 30m keep-alive. CPU-resident coding LLM runs separately and never touches GPU VRAM.
Self-improvement coder — RAM tiers
The Qwen3-Coder-Next model is selected by system RAM, independent of GPU tier. It runs purely on CPU through llama-server and never contends with VRAM.
| System RAM | GGUF Quant | Model Size | Quality | Headroom |
|---|---|---|---|---|
| 56GB+ | UD-Q4_K_XL | ~46GB | Best | ~10GB+ for OS/JARVIS |
| 48-55GB | UD-IQ4_XS | ~38GB | Good | ~10GB+ for OS/JARVIS |
| 32-47GB | UD-IQ2_M | ~25GB | Acceptable | ~7GB+ for OS/JARVIS |
| <32GB | Disabled | would OOM | Do not force-enable | Not enough RAM |
CPU tiers
Strong and beast CPUs can offload ancillary ML from the GPU, freeing VRAM for STT, LLM residency, TTS, and vision.
| CPU Tier | Requirement | Typical Hardware | Effect |
|---|---|---|---|
| weak | <4 threads | SBCs / cheap VPS | Minimal CPU headroom |
| standard | 4-7 threads | Laptop i5 / older desktop | GPU carries ancillary ML when VRAM allows |
| strong | 8-15 threads + 8GB RAM | Desktop i7 / Ryzen 7 | Offloads emotion, speaker ID, embeddings, hemispheres to CPU |
| beast | 16+ threads + 16GB RAM | Ryzen 9 / Threadripper / Xeon | Best partner for premium+ GPUs and coder workflows |
Recommended serious build
Pi 5 sensor body + premium GPU brain.
Premium tier is the sweet spot: qwen3:8b warm, large-v3 STT, GPU TTS, speaker ID, emotion, embeddings, policy, memory, and governance without pretending a bigger LLM is free.