A Show HN post called “I made a Raspberry with Qwen my local car AI” is what sent me down this path: a Raspberry Pi 5 running the open-weight Qwen model family entirely offline, answering questions from a car's owner's manual and reading live vehicle data over voice. You don't need a car or an OBD cable to get the core of it working — you need a Pi, a small Qwen model, and about an hour.
Short answer: You can build a local, offline voice AI on a Raspberry Pi 5 using Ollama and a small Qwen3 model (0.6B–4B parameters) for around $70–$145 in hardware, no subscription required. The full car-integrated version from the original Show HN project uses a much larger 35B model and a 16GB Pi, but the voice-and-answers core runs fine on a cheaper board.
What the original project actually does
The project behind the Show HN post, CarWatch, runs Qwen 3.6-35B-A3B (a mixture-of-experts model with about 3B active parameters) locally through llama.cpp on a 16GB Raspberry Pi 5. According to its README, that setup gets 3.5 tokens per second of generation and 25+ tokens per second of prompt processing, using a 14.3GB quantized build of the model, and it stays around 65°C with active cooling. It listens continuously through whisper.cpp with voice-activity detection (no wake word), answers questions from the car's 745-page owner's manual with page citations, and reports its own system health — temperature, throttling, memory, disk. Commenters on Hacker News were skeptical about some of it, especially whether an LLM can reliably answer things like exact oil capacity across model years, and a few flagged that the README reads like it was written by AI rather than a person. Both are fair points, and worth keeping in mind before you trust any of this with your actual maintenance schedule.
I didn't have a car to wire an OBD reader into, so in my testing I rebuilt the piece that doesn't require one: a Raspberry Pi that listens, thinks locally with Qwen, and talks back. That's also the part most people asking about this project actually want to copy.
What you'll need
A Raspberry Pi 5 is the right board here — earlier Pi models don't have the RAM headroom for even a small language model plus an operating system. Raspberry Pi's own pricing, last updated in a December 2025 announcement, lists the 4GB board at $70, the 8GB board at $95, and the 16GB board at $145, all up from earlier prices because of what the foundation called "the recent unprecedented rise in the cost of LPDDR4 memory," driven partly by AI infrastructure demand. Get the 8GB board at minimum; 16GB only matters if you want to chase the original project's much larger model. You'll also want a USB microphone, a small speaker or the Pi's 3.5mm output, a 27W USB-C power supply, a 32GB+ microSD card or USB SSD, and about an hour where you can leave the board untouched while it downloads and installs.
Step-by-step: a local car AI on a Raspberry Pi
1. Flash the OS and get networked
Use Raspberry Pi Imager to write the 64-bit Raspberry Pi OS Lite (no desktop) to your SD card or SSD — headless and lightweight matters more here than it does on a desktop project. Connect over Wi-Fi or ethernet and confirm SSH works before you do anything else.
2. Install Ollama
Run Ollama's install script over SSH. It's a single command on Linux, and running models locally costs nothing — Ollama's own pricing page confirms the free tier covers unlimited local use; the paid Pro and Team tiers are for its optional cloud-hosted models, not anything you're doing on your own hardware.
3. Pull a small Qwen model sized to your board
Don't reach for the biggest model that technically fits. On an 8GB Pi, start with ollama pull qwen3:1.7b (a 1.4GB download) or qwen3:0.6b (523MB) if you want headroom for the voice pipeline running alongside it. If you've got the 16GB board, qwen3:4b (2.5GB, 256K context) is a real step up in quality and still leaves room to run comfortably.
4. Test it as text first
Run ollama run qwen3:1.7b and ask it a few questions before adding voice to the mix. In my testing, the 1.7B model answered general questions readably at a pace close to what I could comfortably read, while the 0.6B model felt closer to a fast typist — both usable for short answers, neither fast enough for long explanations.
5. Add a voice front end
Install whisper.cpp for speech-to-text and wire a push-to-talk script or a simple voice-activity trigger before you attempt CarWatch's always-listening setup — a manual trigger is far easier to debug when something goes wrong. Pipe whisper.cpp's output into your Ollama prompt, and pipe the model's text response into a basic text-to-speech tool like espeak or piper to close the loop.
6. Run it on a bench before it touches a car
Get the whole loop — mic in, model, speaker out — working reliably on a desk with stable power before you think about 12V adapters, vibration, or mounting anything near a dashboard. CarWatch's more ambitious features, OBD reading and dashcam integration, were still listed as untested on an actual vehicle even in the project's own documentation.
7. Only then consider vehicle integration
If you want to go further, an OBD-to-ethernet adapter and a document store for your car's manual (a plain folder of manual pages works to start) are the next real step. Treat any exact spec the model gives you — oil capacity, torque values, part numbers — as a starting point to verify against your actual manual, not a final answer, exactly as several Hacker News commenters warned.
Example prompts you can copy
These are close to what I ran while testing the text side of this before adding voice:
Summarize what this owner's manual page says about oil type in two sentences: [paste text]— tests whether the model sticks to the source instead of guessing from general training data.I hear a rattling noise when I brake. List three possible causes in order of how common they are.— a reasonable diagnostic-style prompt that a small model can attempt without needing live sensor data.What's the torque spec for a lug nut on a typical passenger car, and why does it matter if I overtighten it?— good for checking whether the model flags "typical" as a caveat rather than stating a false-precise number.Explain what a check engine light for a P0420 code usually means, in plain language.— a common real-world query that doesn't require internet access to answer usefully.Turn this into a short voice reply, under 20 words: [paste a longer answer]— useful once you're piping responses into text-to-speech, since long answers get tedious spoken aloud.
Common mistakes to avoid
I tried the 4B model on an 8GB Pi first and the whole board slowed to a crawl once the voice pipeline was running alongside it — the model alone fit, but model plus whisper.cpp plus the OS didn't, so I dropped to the 1.7B model and it stayed responsive. Second, don't assume a sluggish first run means your hardware is underpowered; Ollama has shipped release candidates with real performance regressions on Raspberry Pi hardware before, so if generation speed drops off a cliff after an update, check for a known issue before you blame the Pi. Third, skip the always-listening voice-activity setup until a push-to-talk version works end to end — debugging a continuously listening pipeline is much harder when you can't tell if the mic, the model, or the speaker is the broken link. Fourth, don't trust a small local model's specific numbers — oil capacity, torque specs, part numbers — without checking them against your actual manual; this is the exact criticism the original project drew on Hacker News, and it's a fair one. Fifth, I underestimated how much a full desktop OS install eats into RAM you need for the model; the Lite, no-desktop image freed up real headroom on the 8GB board.
Tools that make this easier
Ollama is the simplest way to get a Qwen model running locally, and it's what every step above used — free for local use, with paid tiers reserved for its optional cloud models. Whisper.cpp is the standard lightweight choice for offline speech-to-text on a Pi; it's what CarWatch itself uses. If you'd rather skip building the voice pipeline yourself and just want to try Qwen's larger, cloud-hosted models before committing to local hardware, our Qwen vs DeepSeek comparison covers pricing and quality on both, and how to use DeepSeek is worth a look since DeepSeek ships open-weight models that run through Ollama the same way Qwen does.
| Model | Parameters | Download size | Realistic on a Pi 5 with | Notes |
|---|---|---|---|---|
| Qwen3 0.6B | 0.6B | 523MB | 4GB+ board | Fast, best for short commands |
| Qwen3 1.7B | 1.7B | 1.4GB | 8GB board | The sweet spot for a voice assistant |
| Qwen3 4B | 4B | 2.5GB | 16GB board (comfortable) | Noticeably sharper, 256K context |
| Qwen3.6-35B-A3B (CarWatch's model) | 35B (3B active) | 14.3GB quantized | 16GB board only | 3.5 tok/s generation per the project's own numbers |
If your board doesn't clear the RAM bar for any of this, that's a real outcome, not a failure — my AI at home guide covers turning old hardware you already own into a local AI box, and free AI tools rounds up no-hardware options that cost nothing to try instead.
My take
The voice-and-local-model core of this project is genuinely worth a weekend: cheap, offline, and a good way to learn how small a model can be while staying useful. The full car-integrated version — OBD reading, dashcam hooks, a 745-page manual as a knowledge base — is a bigger, harder project than the Show HN title suggests, and the Hacker News skepticism about accuracy on car-specific numbers is warranted. Build the part that talks back first, then decide whether the vehicle integration is worth a second weekend. For picking a model beyond Qwen once you're comfortable with the setup, our best AI models comparison and ChatGPT alternatives for coding guide both cover other open-weight options worth trying on the same hardware.
Frequently Asked Questions
Is building a local car AI with a Raspberry Pi and Qwen free?
The software is free — Ollama, whisper.cpp, and Qwen's open-weight models cost nothing to download and run locally. Your only cost is the hardware: a Raspberry Pi 5 runs $70 (4GB) to $145 (16GB) at current official pricing, plus a microphone, speaker, and power supply.
How long does it take to build a local car AI with a Raspberry Pi and Qwen?
Budget about an hour for the voice-and-model core: 15–20 minutes to flash the OS, 10 minutes to install Ollama and pull a small Qwen model, and the rest for wiring up whisper.cpp and testing prompts. Vehicle integration — OBD, dashcam, a manual knowledge base — is a separate, longer project.
What is the easiest way to do this?
Start with the 1.7B Qwen3 model on an 8GB Pi and test it as plain text over SSH before adding voice. Getting the model answering correctly first, then layering on speech-to-text and text-to-speech, is far easier to debug than building the whole pipeline at once.
Do I need a car to try this project?
No. The voice assistant core — a Pi that listens, thinks with a local Qwen model, and talks back — works fine on a desk with no vehicle involved. Only the OBD and dashcam features from the original CarWatch project actually require a car.
How much RAM do I need for this to work well?
8GB is the realistic minimum for a small Qwen3 model plus a voice pipeline running at the same time. 16GB is only necessary if you want to run something closer to the original project's much larger 35B model.