Run a 27B Qwen Model on Your Laptop for €0 With LM Studio
In short
- Simon Willison ran Qwen 3.8 27B on his own laptop with LM Studio and the llm command line tool. The model file is about 17 GB, and every token is free.
- The price you pay is time: on a Breakout test he needed 52 minutes, while Kimi K3.1 Pro through OpenRouter needed under 10.
- Do the math: an overnight batch of about 10 long jobs costs roughly a euro cent each in electricity. You replace a pay-per-token bill with patience.
Every AI bill is a meter running. You send a long document, the meter ticks, and you pay. For heavy jobs that do not need an answer in ten seconds, there is a cheaper way: run an open model on the machine you already own. Here is how Simon Willison did it this week, what it costs, and where it falls apart.
What the setup is
Three pieces. The model is Qwen 3.8 27B, an open model with 27 billion parameters. He used the Q4_K_M file from lmstudio-community on Hugging Face, which is a compressed version of about 17 GB. LM Studio is the desktop app that loads the model and serves it on your computer. The llm tool, with its llm-lmstudio plugin, lets you call that model from the terminal in one line.
Why did he do it? A research post claimed a stronger AI could read and edit another model’s internal weights. He wanted to see if a small model on a laptop could do the interpretation part too. That is the news hook. The useful part for you is that the same setup runs anything.
What it costs
| Option | Cost per token | Speed in his tests | Your data |
|---|---|---|---|
| Qwen 3.8 27B local | €0 (electricity only) | 46 min Pac-Man, 65 min Pong, 52 min Breakout | Stays on your laptop |
| Kimi K3.1 Pro via OpenRouter | Pay per token, check the current price | 9 min 41 s on Breakout | Sent to a provider |
He does not state his laptop specs, and the article does not name Qwen’s license. Check the license on the model page before you use it in a commercial product. Also check LM Studio’s terms for business use.
Do the math. Say your laptop pulls 60 watts while it works (my assumption, measure yours). The 46-minute Pac-Man run is 0.77 hours, so about 0.046 kWh. At €0.25 per kWh that is roughly €0.01. An 8-hour night fits about ten runs of that length for around €0.12 in total. If the same ten jobs on a paid API would cost you even €1 each, you save about €10 a night. If they cost €5 each, it is €50. You know your own API bill, plug it in.
Step by step
- Check your memory. The file is about 17 GB, so you need clearly more than that free in RAM or GPU memory, plus room for a long context window.
- Install LM Studio from lmstudio.ai and download the Qwen3.8-27B-Q4_K_M.gguf file from the lmstudio-community page on Hugging Face.
- Load the model in LM Studio and set a long context window. The default of 8192 tokens gave Simon errors on long prompts.
- Install the llm command line tool, then add the plugin with
llm install llm-lmstudio. - Raise the timeout, because local reasoning models think for a long time: set the environment variable
LMSTUDIO_TIMEOUT=900. - Run your prompt:
llm -f prompt.txt -m lmstudio/qwen/qwen3.8-27b. The-fflag reads the prompt from a file or a URL. - Walk away. Check the result when it is done.
That is all. No account, no API key, no card.
What to use it for
Pick work that is heavy, repeats, and does not need a fast answer. Summaries of long documents. Classifying hundreds of support emails. Rewriting product descriptions in bulk. Reading a contract you do not want to upload anywhere.
Here is how I would think about it for GiantParts. A catalog of used car parts and ECUs means thousands of near-identical descriptions. That is exactly a night-shift job for a local model, and the stock data never leaves my machine. I have not tested this setup on that catalog, so treat it as a plan, not a result.
Limits and weak spots
Be honest about the speed. On the same Breakout test the local model took about five times longer than the hosted one. Pong took over an hour. If you need an answer while a customer waits, this is the wrong tool.
It also eats your computer. A 17 GB model plus a long context will slow down everything else while it runs. Close your browser tabs, or run it on a second machine.
And a warning on the experiment itself. Simon tested a few puzzles, and one was still running when he published. The model guessed Pac-Man’s game type with 72% confidence, Pong at 75% and Breakout at 88%. That is a fun demo, not a benchmark. Test the model on your own task before you trust it with real work.
A simple test before you commit
Do not switch your whole workflow on a hunch. Take five real prompts from your work, the kind you already pay for. Run each one locally and on your paid tool. Write down three things: how long it took, how good the answer was from 1 to 5, and what the paid run cost.
If the local answers score within one point of the paid ones, move that kind of job to the night shift. If the gap is bigger, keep paying for those and move only the easy tasks. Most people end up with a mix, and that mix is where the saving is.
One more tip: put a reminder in your calendar for a month later and compare your real bill. Savings you do not measure tend to disappear.
Alternatives
- Ollama: another popular way to run open models locally from the terminal.
- llama.cpp: the engine under many local tools, for people who like to control every setting.
- A hosted open model through OpenRouter: faster and still cheap, but your data leaves your machine.
Ivan’s Take
Worth it for night jobs and private data, not worth it when speed matters. I would test it on one real batch task this week: same prompt, local versus paid API, and compare quality and the bill. If the local answers are good enough, you just cut a recurring cost to almost zero.
Sources: Simon Willison · LM Studio
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
