rupayanism

setup 11reel 412026-08-28comment CAR

Run a 35B model on a 300 euro computer (CarWatch setup)

CarWatch runs a 35B model on a 300 euro Raspberry Pi 5 rig, 3.5 tokens per second, no internet needed.

This is the parts list and the setup for CarWatch, the offline car assistant from the reel. Everything below is from the project's own README, checked on 28 August 2026.

What it actually does: a 35 billion parameter model runs on a small computer in the car and answers questions about that specific car, quoting the page number from the owner's manual. No internet, no cloud account, no subscription.

The parts

Part Spec Why this one
Raspberry Pi 5 16GB RAM The 8GB version cannot hold the model. This is the part people get wrong
Active cooling Fan or active cooler Not optional. The README says the chip throttles without it
Storage microSD or NVMe, 32GB or more The model file alone is 14.3GB
Power Official 27W USB-C supply Underpowering a Pi 5 causes throttling that looks like a software fault

Total is about 300 euros. The software is AGPL, so it costs nothing.

The model

Thing Value
Model Qwen3.6-35B-A3B
Quantisation Unsloth UD-Q3_K_S, 14.3GB
Generation speed 3.5 tokens/sec, roughly three and a half words a second
Prompt speed 25+ tokens/sec
Sustained temperature 65 C, with cooling
Knowledge source the car's own 489-page owner's manual, with page citations

Why it fits at all

The A3B in the model name means roughly 3 billion active parameters out of 35 billion total. Only a fraction of the model fires per token. That is the whole reason a 14.3GB file runs on a board this size.

What to expect, honestly

It is slow. Three and a half words a second means a paragraph takes most of a minute. If you are expecting ChatGPT's speed you will hate it.

The constraint is the product. It is not trying to be a general assistant. It answers questions about one car, from one manual, and that narrowness is why a model this size is useful on hardware this small.

Cooling is a hard requirement. Skip it and the chip throttles, and you will blame the model.

Scanned manuals need OCR first. The retrieval works on text. If your manual is a scan, that is a separate step before any of this.

The source

The project: https://github.com/ThinkOffApp/CarWatch

Every number on this page is from that README. If it has changed since 28 August 2026, trust the repo over this page.

Which car are you building it for?

Tell me in The Rabbit Hole and I will tell you whether your manual needs OCR first.

It is the room where I keep the parts that don't fit on a page like this: the full run of numbers, what broke the first time, and the ones that looked good and then didn't work at all.

Two catches, up front, because that's the rule here. It's private, so you have to be inside to read anything. And it's new, so you'd be early rather than joining something busy.

Free: skool.com/the-rabbit-hole-3206/about

if you run it, tell me

the rabbit hole

It's free to join, and the course library is free too. Ask there and I answer in the open, so the next person with the same question finds it. There's an optional paid Premium tier for weekly office hours.

join the rabbit hole

Post in there what you ran it on and what you got, or reply to the DM. I read every one.