THIS RETRO TALKING HEAD RUNS A PRIVATE CHATGPT-STYLE AI FULLY OFFLINE ON ONE RASPBERRY PI
A talking Max Headroom head that runs a full voice LLM locally — no cloud, no API calls, and it can write and run programs for itself.
AIOpen-hardware
- difficulty
- ●●●●○
- time
- a weekend-plus
- license
- GPL-3.0
- repo
- repo ACTIVE349 stars
●●●●○ · a weekend-plus · GPL-3.0 · 349 stars · repo ACTIVE
WHAT YOU’LL NEED
- 3D printer + filament — printable parts — files are in the repo
- Electronic parts — full list with part numbers in the repo BOM
- Dev board / microcontroller — runs the project firmware
Partner
COMPAREE VERDICT
Max Headbox is what happens when someone who actually uses LLMs daily decides the cloud dependency is unacceptable. It runs Whisper for speech recognition, Ollama with Llama models for conversation, and Piper for text-to-speech — all on a Raspberry Pi 5 with no internet required after setup. The head servo moves with speech for that Max Headroom stutter effect. The genuinely novel piece is the function-calling layer: you can give it a task like 'remind me in 10 minutes' and it will write a small Python script, show it to you, and run it if you approve. That is not a party trick — it is a working offline agent architecture. The catch: inference is slow (10-30 seconds per response depending on model size), you will spend more time than expected tuning Whisper's wake-word sensitivity to avoid false triggers, and the servo mount screws are fiddly enough that you will wish there were alignment jigs in the STL pack. If you have run Ollama before and understand what 'quantized model' means, this is a weekend build. If those words are new, budget an extra evening for the learning curve. The real limitation is not difficulty — it is whether you can live with response times measured in double-digit seconds. For a private AI that never phones home, that trade-off makes sense. For anything needing snappy replies, it does not.
IN THE REPO
NOT IN THE REPO
- —Full STLs for the talking head enclosure, detailed bill of materials with part links, complete software stack in the repo.
- —Installation docs cover Raspberry Pi setup, Whisper model download, Ollama configuration, and servo wiring.
- —No PCB — the build uses off-the-shelf servo controllers and a USB mic.
- —GPL-3.0 license permits commercial use with source disclosure.
- —The creator built this as a learning project and documents the entire reasoning behind model choices.
- —Active development — last commit within the past month, issues are answered.
Can I build this?
Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.
Partner · KickstarterHeyGears G1: 10M+ colors and transparent parts in one print, plus UV printing on flat objects. Figures, parts and labels — no painting.
Videos
Max Headbox an LLM agent that runs on a Raspberry Pi
No official project video or channel URL was provided in the payload.
Gallery
Start here
Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.
- 1.Read the architecture overview (Understand the Whisper → Ollama → Piper pipeline and what each component does before installing anything.)
- 2.Print the enclosure(5 parts, PLA at 0.2mm layer height. The jaw hinge needs supports; everything else prints clean.)
- 3.Gather the BOM (Linked parts are verified. Do not substitute the servo without checking torque spec — the jaw needs at least 1.5 kg·cm.)
- 4.Install the software stack (Scripted install handles Whisper, Ollama, and Piper dependencies. Budget 1-2 hours for model downloads on a decent connection.)
Resources
Documentation, files and community threads for this build — we link straight to the original sources and never rehost the creator’s files.
Partner · KickstarterCORE POWER BANK S: 0.2" thin, 2.8 oz, 5000 mAh with 22.5W fast charging. Semi-solid battery, carbon fiber, IPX7 waterproof.
KNOWN ISSUES
- The Pi 5 4GB model works but swaps heavily under load — 8GB is the practical minimum for smooth inference with anything larger than Llama 3B.
- Whisper's wake-word detection has false positives if threshold is too low, misses you if too high. Expect to spend 30+ minutes in the config file finding the sweet spot for your room noise.
- The servo bracket screws into plastic bosses with no alignment marks — dry-fit first or you will strip the first hole learning the angle.
- Ollama models are large (2-7GB) and you need at least one downloaded before the voice agent works. The README assumes you know which quantization level to pick; if not, start with the 4-bit q4_0 variant.
- Response latency is not a bug — inference on a Pi takes 10-30 seconds depending on prompt length and model size. If you need sub-second replies, this hardware cannot deliver it.
- Function-calling (the 'write a program' feature) requires approval mode enabled or it will execute arbitrary code. The docs explain this but it is easy to miss in the install script.
Can I use a Raspberry Pi 4?
Technically yes, but inference will be painfully slow — expect 45+ seconds per response. The Pi 5's faster CPU and RAM bandwidth are what make this remotely practical.
Does it work in languages other than English?
Whisper supports 50+ languages, Piper TTS has multi-language models, and you can run non-English Llama variants through Ollama. The creator tested Italian; check the Whisper and Piper model lists for your language before committing.
How much does response time improve with a bigger model?
It does not — bigger models are slower. Llama 8B takes nearly twice as long as 3B. You trade speed for coherence, and on a Pi the speed cost is brutal.
Can I add a camera or make it control smart home devices?
The function-calling architecture is designed for exactly that — you write a Python function, expose it to the agent, and it can call it during conversation. The README has an example for setting reminders; extending to camera capture or API calls follows the same pattern.
Community builds
No community builds yet — be the first, we feature the best ones.
Discussion1
FROM THE COMPAREE TEAM
348 stars and the entire stack runs on a single Pi with zero cloud dependency — inference takes 10-30 seconds per reply, but nothing ever leaves the box. Would you trade response speed for that level of privacy?
Simone (syxanash)
Simone built Max Headbox to prove a voice AI agent could run entirely offline without sacrificing functionality. The project doubles as a learning repository for anyone wanting to understand how modern LLM pipelines work under the hood, with clear reasoning in the docs for every model and architecture choice.
Keep building — related projects
Same category, picked by how close they are to this build.
NO CAD FILESBuild a talking AI assistant on a breadboard and host its brain yourself
AIOpen-hardware●●●○○moderate
FULL CADTurn any glasses into AI smart glasses for 20 dollars - and it remembers everyone you meet
AIOpen-hardware●●●○○moderate
NO CAD FILESA 28.9M parameter language model runs entirely on an eight dollar microcontroller
AIOpen-hardware●●●○○moderate
NO CAD FILESA 28.9M parameter language model runs entirely on an 8 dollars microcontroller
AIOpen-hardware●●●●○advanced
DISCLAIMER
- Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
- Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
- Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.
CompareeTEAM25d agoedited
Practical notes from our verification: the repository is actively maintained with the last commit in mid-July 2026, there is no official video walkthrough (the reel footage comes from the README images), and the single biggest time sink is not assembly — it is tuning Whisper's wake-word threshold in a real room with ambient noise. The creator is transparent about inference speed: this is not a snappy assistant, it is a working proof that offline voice agents are viable if you can stomach the latency. The function-calling layer is the genuinely novel piece — it is not a demo, it is a real architecture for extensibility.