DeepSeek, Mistral and AI that runs on your own computer
In Italy the word «deepseek» is searched on Google about 301,000 times a month. «Ollama», the name of a program that runs artificial intelligence directly on your computer, reaches 22,200 searches, more than double a year ago, and «lm studio», a similar program, 12,100. These are Google Keyword Planner figures for the last year, and in my view they reveal a precise curiosity. Many people have heard that there are «open» AIs you can use without sending your texts to a company, and they want to know whether it is true.
It is true, but with some conditions worth knowing before you install anything. In this article I try to put them in order, starting from the most searched case, DeepSeek, which is also the one that causes the most confusion.
An app and a model are not the same thing
When people say «DeepSeek» they mean two different things.
The first is the service, that is the website and the app where you type your questions. Its privacy policy, updated on 10 February 2026, states explicitly that personal data is collected, processed and stored in the People's Republic of China. On 30 January 2025 the Italian data protection authority (Garante privacy) ordered a restriction on the processing of Italian users' data, without prejudice to any other decision at the end of its investigation, and the Garante's website shows no later decision. On 21 October 2026 the app was not available in the Italian App Store, while it was in Germany, even though in June 2025 the Berlin Commissioner for Data Protection had notified Apple and Google of it as illegal content, because of the transfer of data to China. In Italy the competition authority (AGCM) also looked into it, but for another reason, namely the warning to users about made-up answers (so-called hallucinations), and it closed the proceedings in December 2025 by accepting the company's commitments.
The second thing is the model, that is the file with the parameters the AI learned during training, called weights. DeepSeek publishes them on Hugging Face under the MIT licence, one of the most permissive there is. Anyone can download them and run them on their own servers, and in that case the texts do not go through the Chinese company. There is, however, a detail that changes everything. The latest model, released on 17 October 2026, has about 763 billion parameters, and a laptop cannot run it (we will see the numbers shortly).
That is why when someone tells you «DeepSeek is open, so it is safe» they are only half right. The model is open, but the app you probably have in mind is not, and the full model needs data-centre machines.
Open, but how open?
The word «open» also deserves attention, because in AI it is used with different meanings.
In most cases people mean open weights, that is you can download the model and use it, under a licence that varies from company to company. A fully open source model is something different. According to the definition published by the Open Source Initiative, the organisation that since 1998 has set what counts as open source in software, three things are needed together, the weights, the complete code used to train and run the model, and information about the training data detailed enough to rebuild an equivalent one. Few famous models meet them all.
The licences of the most searched models, in brief.
- DeepSeek publishes its recent models under the MIT licence, which also allows commercial use.
- Mistral AI, the French company that created Le Chat (now Vibe), publishes part of its models under the Apache 2.0 licence, including the Ministral 3 family, Mistral Small 4 and Mistral Large 3. This does not apply to all of them, because some have modified licences, one is restricted to non-commercial use and others remain proprietary.
- Meta's Llama uses its own licence that the Open Source Initiative does not consider open source. For the multimodal Llama 4 models, the ones that also understand images, the licence excludes anyone domiciled or with a principal place of business in the European Union, while people using a product that contains them are not affected by this exclusion.
The practical conclusion, in my view, is to always read the licence of the model you download, especially if you plan to use it for work.
How much memory you need to run it at home
Here a simple rule comes into play. Hugging Face, the platform where many of these models are published, explains that loading a model in 16-bit precision takes about 2 GB of memory for every billion parameters. Programs for local use often use versions compressed to 4 bits, which take a quarter of that space. From this I derive, and it is my own calculation, that you need about 0.5 GB per billion parameters, plus a margin for the conversation in progress.

Estimated memory for the weights only, compressed to 4 bits. Source, my calculation from the sizes published on Hugging Face and the rule in the Hugging Face documentation, checked on 21 October 2026.
The chart says something advertising does not. Small models, from 3 to 14 billion parameters, fit on a recent computer with 16 GB of RAM, which is the amount recommended by LM Studio. The large models you read about in the news do not, they need hundreds of gigabytes. So locally you will use a smaller model, which does well on simple questions, summaries and rewrites, but makes mistakes more often than an online service on complex tasks.
Ollama, LM Studio and GPT4All compared
To use a model on your computer you need a program that downloads it and runs it. The three most searched in Italy among those I measured are different from one another, and it is worth knowing how.

Three programs for using AI on your computer. Source, the programs' official websites and repositories, checked on 21 October 2026.
Ollama is free and its code is open under the MIT licence. It runs on macOS, Windows and Linux and since July 2025 it also has an app with windows and buttons, as well as the terminal. Its documentation says that when you use it locally the company does not see your prompts or your data. There is, however, an important caveat. Ollama also offers cloud models, recognisable by the «:cloud» tag in their name, and in that case requests are processed on Ollama's servers. If you want to be sure you stay local, the documentation explains how to switch off the cloud features, with the variable OLLAMA_NO_CLOUD=1 or with the setting "disable_ollama_cloud": true in the file ~/.ollama/server.json.
LM Studio is the easiest to use for people who do not like the terminal. Since 8 July 2025 it has been free for work too, but its code is not open. On the Mac it requires an Apple Silicon processor and macOS 14 or later, on Windows a processor with AVX2 instructions, and in both cases it recommends 16 GB of RAM.
GPT4All is open under the MIT licence and also works without a graphics card, using the processor alone. Its latest version, however, dates from February 2025, so today I consider it less suitable than Ollama and LM Studio.
What changes for privacy, and what does not
With a local model your questions do not travel to a company, and for anyone who works with sensitive data that is a huge advantage. A few things still need to be known.
- Download models only from official sources, such as the companies' pages on Hugging Face or the libraries built into the programs.
- Check the features that use the internet, such as cloud models or web search, because there your texts leave the computer.
- Double-check the answers, because a small model gets things wrong more often.
- Personal data is still personal data. If at school or in a professional firm you handle other people's data, even with a local tool the decisions should be made with your organisation's data protection officer.
The same principle, doing the sensitive work on your computer before anything goes online, is the one behind the Anonimizzatore IA (AI Anonymiser) that we distribute free of charge for macOS and Windows. Personal data is recognised on your computer, your documents stay there, and only afterwards can you paste the text with placeholders into any chatbot, online ones included. Always check the result, because no tool recognises everything. If you want to understand how placeholders work, I wrote about it in the guide on how to anonymise texts for chatbots, and if you use the free online versions you will find the comparison in the article on free AI chatbots and your data.
If you want to receive updates like this one, and the free licence for the Anonimizzatore IA, create your free account on retoria.it (the form is in Italian). If instead you want to try local AI with colleagues at your school or office, guided step by step, write to us.
Methodological note
On 21 October 2026 I read the official pages of the services and programs (privacy policies, documentation, licences, model cards on Hugging Face) and the authorities' decisions. I checked the availability of the DeepSeek app in the Italian and German App Stores that day with Apple's official lookup service. For Google Play I do not have a verified figure.
The memory in the chart is my own estimate, obtained by multiplying the parameters published on Hugging Face by 0.5 GB per billion, that is a quarter of the 2 GB per billion indicated by Hugging Face for 16-bit precision. It covers the weights only, and in practice you need an extra margin that grows with the length of the conversation. The parameter count of Mistral Large 3 is the one given in the model's name.
I found no official Italian guidance (Garante, AgID, Ministry of Education) recommending the use of local models. Search volumes come from Google Keyword Planner for Italy, from October 2025 to September 2026. This article is not legal advice.
Sources
- DeepSeek, privacy policy (updated 10 February 2026) and models on Hugging Face, with the DeepSeek V4.1 Flash model card (17 October 2026, MIT licence).
- Italian data protection authority (Garante per la protezione dei dati personali), restriction order against DeepSeek (30 January 2025).
- Berlin Commissioner for Data Protection and Freedom of Information, press release on DeepSeek (27 June 2025).
- Italian Competition Authority (AGCM), press release on closing proceedings PS12942 with commitments (December 2025).
- Mistral AI, models overview, Hugging Face cards for Ministral 3 8B, Mistral Small 4 and Mistral Large 3, and page on the move from Le Chat to Vibe (12 August 2026).
- Meta, Llama 4 acceptable use policy, and Open Source Initiative, article on the Llama licence.
- Open Source Initiative, Open Source AI Definition 1.0.
- Hugging Face, documentation on optimising language models.
- Ollama, FAQ, page on cloud models and announcement of the new app (30 July 2025).
- LM Studio, announcement of free use for work (8 July 2025), terms of use (23 August 2026) and system requirements.
- Nomic AI, GPT4All page and GitHub repository.
- Google, Keyword Planner, average monthly search volumes in Italy (October 2025, September 2026), checked on 21 October 2026.