We Ran a Local AI Model on a Six Year Old Budget Laptop Chip. Here Is Exactly What It Could Do.

Questo articolo non è ancora disponibile in italiano. Viene mostrata la versione inglese.

The verdict box

• What it is: running an AI model on your own computer, with no account, no subscription and no internet connection
• What we tested: Qwen2.5 1.5B Instruct, a small open-weight model, in Q4_K_M form, a compressed version about 1.1 gigabytes in size
• Who makes it: the model comes from Alibaba, the engine from the open source llama.cpp project
• On what: an AMD Ryzen 3 3250U, a low end laptop processor from 2020, two cores available, about 3.8 gigabytes of memory, no dedicated graphics card
• Tested: 25 July 2026, llama.cpp build b10107
• Cost: 0 euros, no account
• Available in the EU: yes, nothing to sign up for and nothing to check
• Speed: 4.8 to 6.9 tokens per second of output, which is roughly 4 words per second
• Best for: reshaping text you already have, offline, privately
• Not for: anything where the facts have to be right
• In one sentence: on hardware this modest, local AI is a text shaping tool rather than a knowledge tool, and within that limit it is genuinely useful

---

Most guides to running AI on your own computer start with a graphics card that costs more than a used car. That is fair for the big models. It also quietly tells millions of people with an ordinary laptop that this is not for them.

So we went the other way and asked the opposite question. Not “what is the best local setup”, but what is the floor? What happens on a machine nobody would call a workstation.

The answer is more interesting than we expected.

What “running it locally” actually means

When you use ChatGPT or Claude, your text travels to a data centre, a model there produces an answer, and it comes back. Running a model locally means the model file sits on your own hard drive and your computer does the thinking. Nothing leaves the machine. It works on a train with no signal.

The price is size. A model you download has to fit in your memory, so the ones you can run at home are much smaller than the ones in data centres. Smaller means it knows less and reasons less well. The whole question is whether what remains is still worth having.

Three things make this possible on modest hardware. Quantisation compresses the model, roughly like saving a photo as a smaller JPEG: some detail is lost, the file gets much smaller. Small models have improved enormously, and a model with 1.5 billion parameters, the adjustable numbers a model learns during training, now does things that needed ten times the size two years ago. And llama.cpp, the open source engine we used, is built to run on ordinary processors rather than expensive graphics cards. That engine sits underneath the friendly apps most people use, Ollama, LM Studio and Jan.

How we tested

We ran llama-cli from build b10107 on a Linux container, calling the file qwen2.5-1.5b-instruct-q4_k_m.gguf with two processor threads, a 512 token context window, meaning how much text the model can hold in mind at once, and a temperature between 0.2 and 0.3, which makes answers more predictable and less inventive. The model came from Hugging Face at about 1.1 gigabytes.

Then we ran four tasks from the set we put every tool through, and wrote down what came back, including the parts that did not work.

Being straight about what we did not test. We drove the engine from the command line, not through Ollama, LM Studio or Jan. Those apps wrap the same engine in a friendlier front end, so the speeds should be in the same range, but we did not measure them and will not claim we did. We skipped three of our standard tasks: long documents and multi step jobs need more context than this machine can hold, and we did not test a second European language beyond German. Each prompt was run once, so treat single results as indicative rather than averaged. And one thing about the model itself: Qwen2.5 was released in September 2024. We picked it because it is the small model most widely mirrored and best documented, not because it is the newest. Newer small models have later knowledge cut offs, so the speed figures below should carry over, but the complaint about stale knowledge may not.

The measurements

Speed first, because it decides whether any of this is usable.

• Reading your input: 11.3 to 12.5 tokens per second
• Writing its answer: 4.8 to 6.9 tokens per second
• Loading the model from disk: roughly 20 seconds the first time, quicker afterwards when the file is still held in memory

A token is roughly three quarters of a word, so the output works out to about 3.5 to 5 words per second. That is close to a slow reading pace. You watch the words appear one after another, and for a short answer you wait ten to twenty seconds.

Memory sat comfortably. The model occupied around a gigabyte, the machine had nearly four, and nothing had to be shuffled out to disk.

What it did well

Structuring messy text was flawless. We gave it a deliberately sloppy shopping list: milk 1.29 two bottles; bread three euro fifty one loaf; 6 eggs 2,80 EUR, and asked for a table with columns for item, price and quantity. This came back:

| Item  | Price (EUR) | Quantity |
|-------|-------------|----------|
| Milk  | 1.29        | 2        |
| Bread | 3.50        | 1        |
| Eggs  | 2.80        | 6        |

It converted “three euro fifty” into 3.50, read the comma decimal in “2,80” correctly, and understood that “6 eggs” meant a quantity of six rather than a price. On a six year old budget chip, in a few seconds.

This is the category that works. Reformatting, extracting, tidying, converting. Text goes in, better organised text comes out, and the model needs no knowledge of the world to do it.

It knew what it did not know. Asked who the current German chancellor is and when the EU AI Act transparency rules took effect, it answered: “As of my last update in 2023, the current Chancellor of Germany is Olaf Scholz. However, I don’t have specific information on when the EU AI Act transparency rules took effect.” The first part is out of date. The second part is the good bit. A small model volunteering the limits of its own knowledge rather than inventing a date is better behaviour than we have seen from far larger models.

Drafting worked, at a basic level. A polite English email declining a meeting came back structured, correctly formatted, with sensible placeholders in square brackets for the name and the reason.

Where it fell down

German was noticeably weaker than English. Asked for a polite decline in German, it produced something understandable but stilted, including “Absagezeilen”, which is not a word anyone uses, and a phrase about committing oneself to an important project that reads like a translation. For our readers this matters more than any benchmark: if you work in German, French or Italian, a small local model will feel clumsier than the same model in English. Generation also slowed to 4.8 tokens per second on the German task, the slowest of the four.

The drafts are generic. The email was correct and forgettable. It cannot know your tone, your situation, or why you are really declining. On a paid cloud assistant with memory, that gap closes. Here it does not.

Its knowledge is stale and you cannot fix that. The model’s world stopped in 2023. There is no web search, no updates, no memory. For anything current, this is the wrong tool and no amount of prompting changes it.

Waiting is part of the experience. Twenty seconds to load, then five words a second. For a quick reformat that is fine. For a long piece of writing you will feel every second.

What it costs, and getting it in Europe

Nothing, which is the headline. No subscription, no account, no credit card, no waiting list. The real price is 1.1 gigabytes of disk space, about a gigabyte of memory while it runs, and your patience.

The model we used, Qwen2.5 1.5B Instruct, is published by Alibaba under the Apache 2.0 licence, which permits commercial use. That is not true of every open model, so read the licence before you build a business on one. llama.cpp is MIT licensed. Neither is geographically restricted, so there is no European availability problem and no feature that arrives here six months late.

On privacy, this is the strongest position available. Your text never leaves the computer. There is no provider to send a data request to, no terms of service to read, no training opt out to hunt for in a settings menu, because there is no transfer at all. For anyone handling client data, medical notes or anything a European data protection officer would want documented, the difference between “processed in the EU” and “never left this laptop” is not a small one.

Who it is for, who should skip it

Worth doing if you have an older computer sitting around, you handle text you would rather not upload, you work offline sometimes, or you want to understand how this technology actually works by holding a piece of it. Also if you simply object to paying a monthly fee for something a machine you already own can do.

Skip it if you need facts, current information, or polished writing in a language other than English. Skip it if waiting annoys you. And skip it if you were hoping to replace a paid assistant, because at this size you cannot, and pretending otherwise would waste your evening.

Alternatives

Ollama is the friendliest way in for most people. One installer, one command, and you have a model running. It is the same engine underneath, with a library of models you can pull by name.

LM Studio and Jan give you a window with a chat box instead of a terminal, model browsing built in, and no command line at all. If the idea of typing commands puts you off, start with one of these.

A bigger model on the same idea. If your machine has 16 gigabytes of memory or a decent graphics card, models in the 7 to 12 billion parameter range are a different experience: better in German, better at reasoning, still free and still private. Gemma 4 12B is a good example of how far this class has come.

A European cloud model, if local is not enough. Mistral processes in the EU by default, as of July 2026, which is the next best answer after “it never left my laptop”.

The verdict

We expected a machine like this to fail, and it did not. It failed at some things, specifically at knowing anything and at writing decent German, and it succeeded completely at the thing most people actually need day to day: taking messy text and making it tidy.

If that sounds like a small win, consider what it costs. Nothing, forever, offline, on hardware you were about to throw away.

The honest framing is this. A small local model is not a smaller version of ChatGPT. It is a different tool that happens to share an interface. Judge it as a text processor with a language understanding module attached, and it is impressive. Judge it as an assistant, and it will disappoint you within five minutes.

Tested on 25 July 2026 with llama.cpp build b10107 and Qwen2.5 1.5B Instruct Q4_K_M. Local AI is moving quickly and small models improve every few months. Check the date on this piece before you trust the numbers.

Sources

• llama.cpp releases (build b10107 used for this test)
• Qwen2.5 1.5B Instruct GGUF model card, Hugging Face
• Ollama
• LM Studio
• Jan