Show HN: I made a Raspberry with Qwen my local car AI
This project showcases an audacious feat: running a 35B-parameter Qwen model locally on a Raspberry Pi 5 to turn a car into an intelligent, offline chat-room agent. It integrates OBD data, leverages the car's manual for RAG, and connects to other AI agents. The Hacker News crowd is both impressed by the technical ambition and skeptical about its real-world practicality and claimed 'proven' status.
The Lowdown
The CarWatch project by ThinkOffApp transforms a vehicle into a "chat-room agent" using a Raspberry Pi 5. By running a large language model (Qwen3.6-35B-A3B) entirely offline on the Pi, it aims to provide car-specific insights and control, interfacing with the vehicle's diagnostics and even manufacturer cloud services.
- Local AI Powerhouse: A Raspberry Pi 5 (16GB) hosts a Qwen3.6-35B-A3B model, achieving impressive generation speeds (3.5 tok/s) and prompt processing (25+ tok/s) without relying on cloud services.
- Grounded Knowledge: The system incorporates the car's full owner's manual (745 pages) using lexical RAG, ensuring answers are grounded and cited, and explicitly refuses to answer beyond its known data.
- Autonomous & Voice-Enabled: It features a continuous, hands-free voice interface (VAD + whisper.cpp on-Pi), self-starts all services on boot, and pulls updates from GitHub.
- Deep Integration: It connects to the car's OBD system for real-time data, and aims to interface with manufacturer cloud services for controls like AC or door locks.
- Connectivity Strategy: Prioritizes local, offline functionality for critical safety information and gracefully queues social or heavy tasks for when connectivity is available.
- Honesty Policy: The project emphasizes a rigorous
The Gossip
Plausibility Probed
Many commenters express significant skepticism regarding the project's practical claims, particularly about successful OBD integration and interfacing with proprietary manufacturer cloud services. They question how the system establishes TLS connections to car APIs and whether certain features, like OBD engine reading, are truly 'proven' on a real car rather than just 'built + tested' against simulations. Some even suggest the project serves primarily to promote the author's other services like GroupMind.
Processing Power Ponderings
The choice of a 35B Qwen model on a Raspberry Pi 5 sparked a technical debate. Commenters discussed the feasibility of running such a large model given the Pi's memory bandwidth limitations, with some suggesting smaller, more optimized models (like Gemma 4B/9B or Qwen 3.5-4B) would be more practical and efficient. There's an underlying question of whether the computational power is overkill for the stated use cases.
Automotive Answers Appraised
Users raised concerns about the accuracy and utility of LLMs for specific automotive information, noting that models can struggle with subtle details like specific oil types for different year/model variations. While the RAG approach using the owner's manual is lauded as a good strategy, the overall necessity and 'point' of having an LLM to answer manual questions, or as a full 'chat-room agent' for a car, is a point of contention. Some argue that simpler search or direct lookup tools might be more efficient for many of the proposed functionalities.