A local AI model for your business: Gemma 4, hardware and costs

An accounting office doesn’t want to paste its clients’ invoices into ChatGPT. A law firm prefers not to send contracts to the cloud, and an online store would like to write hundreds of product descriptions without paying for every request. In each of these situations it is worth considering a local AI model for your business: artificial intelligence that runs on your own computer or server and does not send your content to a provider.

I show what Google’s Gemma 4 can do, what hardware is enough and how much it costs compared with the cloud. I describe the state of things at the beginning of November 2026.

What a local AI model is and what Gemma 4 can do

ChatGPT and Gemini run on the providers’ servers. You download a local model once, and after that your computer does all the computing. The content of an email, contract or invoice never leaves the company, and you don’t pay for requests.

Free programs are enough. The simplest is Ollama (Windows, macOS, Linux): you install it, type one command and talk to the model. For a team, you add a browser-based chat such as Open WebUI, which works without the internet and supports accounts for many people as well as knowledge bases built from documents.

Gemma 4 is a family of open Google models released in early April 2026. It is the first Gemma under the Apache 2.0 license, which lets you use the model in a business with no license fees. There are five sizes:

• E2B and E4B: small models for laptops and phones (“E” stands for effective 2 and 4 billion parameters),
• 12B: a mid-size model for a computer with a graphics card,
• 26B A4B: an MoE-type model that uses about 4 billion parameters for each word, so it usually answers faster than 31B, but it keeps the whole thing in memory,
• 31B: the largest model, in which all the parameters work.

According to Google’s documentation, all sizes understand text and images, and E2B, E4B and 12B also understand audio. Small models read up to 128 thousand tokens at a time, larger ones up to 256 thousand, so they fit long contracts. The models can use connected tools and have a thinking mode for harder tasks. The model card states that Gemma 4 was trained on more than 140 languages and supports more than 35 “out of the box”. So check the model’s Polish on your own texts.

A Polish alternative: Bielik

Bielik is a Polish model developed by the SpeakLeash Foundation and ACK Cyfronet AGH (the Academic Computer Centre Cyfronet of AGH University of Krakow). The newest line is version 3, e.g. Bielik 11B v3.0 and the smaller Minitron 7B v3.0. It has an Apache 2.0 license and is geared toward Polish, but it works on text only, so Gemma will be better for images and audio. Details: bielik.ai. It is worth testing both models on the same tasks.

What local AI is useful for in a business and a store

A local model works best for repetitive work with text when data should not leave the company:

• Team chat: drafts of letters, translations and text editing, only on your own hardware.
• Answers from company documents: the model searches terms and conditions, procedures and price lists.
• Product descriptions: hundreds of descriptions from manufacturer data with no per-request fees.
• Email drafts and replies to complaints: the model writes a working version, a person checks it.
• Summaries of contracts and notes: deadlines, amounts and obligations in a few points.
• Reading invoices and photos: data from an invoice into a table, a description of a product photo.

A prompt for testing:

Prepare a draft reply to the complaint below. Polite and factual tone, 120 words at most. Do not promise a refund before the complaint is reviewed. At the end, list what we still need to check. [complaint text]

When the model is meant to act on its own, e.g. check stock levels or create tasks in a CRM, that is already process automation.

Local AI standing guard over security: it reads logs and warns you

Your server, website, email and firewall record every login and suspicious request in logs. In a small company nobody reads these files every day. A local model can do it: every hour or once a day it gets a fresh chunk of logs and says in plain language what looks suspicious, for example a series of failed logins to one account or scanning hundreds of website addresses within a few minutes. The warning goes to email or a messenger, and logs with IP addresses and usernames never leave for the cloud.

An example prompt:

Below are the server login logs from the last 24 hours. List suspicious events in plain language: series of failed logins, logins at unusual times, scanning of many addresses. For each, give the time, the IP address and the risk: low, medium or high. Do not add events that are not in the logs.

An honest word about the limits: this is an extra pair of eyes, not a security system. The model will not replace a specialist, updates, backups or tools that block attacks on their own (e.g. fail2ban). It may miss an attack or raise a false alarm, and large logs have to be filtered by a script first. Logs also contain text typed by an attacker, so the model only reports and gets no permissions to change anything on the server. I do server hardening, monitoring and a failure plan in the Project cybersecurity service.

Which hardware: three tiers

The most important thing is memory: of the graphics card (VRAM) or the shared memory in Macs and mini PCs with AMD Ryzen AI processors. Google states in the Gemma documentation how much memory the model weights alone need, with a 20% margin:

The table does not include context, meaning the text the model reads. The longer the document and the more people ask at the same time, the more memory is needed on top of these numbers. Hence three tiers:

1. To try it out: a computer you already have. A laptop with 16 GB of memory will run E2B and E4B even without a graphics card, just more slowly. That is enough for testing and for one person’s work.
2. For a small team: a card with 16 GB of VRAM. The 12B model fits with room for longer documents, so it will handle a chat for a few people and product descriptions.
3. A powerful company server: 32 GB of VRAM or a lot of shared memory. An RTX 5090 card (32 GB) or a mini PC with an AMD Ryzen AI Max+ 395 and 128 GB of memory will run 26B A4B and 31B with a long context.

Response speed depends on the hardware, the model version and the settings, so check it on your own task before buying.

What it costs and when the cloud is cheaper

In 2026 the prices of cards and memory rose sharply and change often, so treat these amounts as approximate (retail prices):

• RTX 5060 Ti 16 GB: about PLN 3,000 according to ppe.pl from August 2026,
• RTX 5090 32 GB: about PLN 30,000 in Polish stores according to benchmark.pl from 20 September 2026, against NVIDIA’s suggested price of PLN 9,100,
• mini PC with a Ryzen AI Max+ 395 and 128 GB: about €3,900 in the EU in February 2026.

My Local AI model for your business service costs from PLN 2,990 net plus hardware: I choose the hardware, install the model and test it on your tasks. Support costs from PLN 199 net a month.

In the guide ChatGPT or Gemini for business I counted about PLN 100 per person a month for a business cloud plan, excluding VAT. Here is an example for 5 people, where the card goes into a computer you already have. All amounts are net: a VAT-registered company deducts the tax on the card, so PLN 3,000 in the store is about PLN 2,440 net.

After the first year the local model works out cheaper, and additional people don’t raise the cost as long as the hardware is sufficient.

A local model makes no sense when:

• the tasks are the hardest: for complex analyses the most powerful cloud models are usually better,
• you use AI rarely: with a few questions a week a cloud plan will be cheaper,
• nobody will look after it: updates, backups and a computer failure are your responsibility.

Often the best option is a combination: a local model for sensitive data, the cloud for the rest. I work out such a split in the AI implementation in your business service.

GDPR and the AI Act: the data stays in the company, and so do the obligations

With a local model your content does not go to a provider, so the question of training on your data and transferring it outside the EU goes away. The GDPR still applies, though:

• decide who has access to the chat and the document base (separate accounts, not one shared password),
• decide how long you keep conversation history,
• secure the computer with the model: disk encryption, backups, updates,
• if an external company has access to the server, for example as part of support, sign a data processing agreement.

Watch out for a trap: Ollama models with the cloud suffix (e.g. gemma4:cloud) run on Ollama’s servers, so the content leaves the company. Turn off cloud features in the settings. I set up a separate, encrypted server for customer data in Poland or the EU in the Customer data in the EU service.

When the model talks to customers, for example as an AI chatbot on your website, from 2 August 2026 Article 50 of the AI Act requires that the customer know they are talking to AI. I will adapt the chat as part of the AI Act package. This is general information, not legal advice.

Rollout step by step: test first, buy later

1. Pick 5–10 real tasks: a complaint email, a contract to summarize, products to describe, an invoice scan.
2. Run a small model on a computer you already have. Install Ollama and type in the terminal:

ollama run gemma4:e4b

The first run will download the model (a few GB). On a more powerful computer, also try gemma4:12b. Google has the instructions in the Gemma documentation.

3. Compare with the cloud. Give the same tasks, without personal data, to a cloud model. Assess the Polish, the errors and the waiting time.
4. Choose hardware for the model that passed the test, not the other way round. If E4B is enough, don’t buy an expensive card.
5. Set up a chat for the team: accounts, a knowledge base from documents, backups.
6. Write down the rules. The model makes mistakes, so a person always checks numbers, deadlines and replies to customers.

Summary

A local AI model makes sense when you work with data that should not go to the cloud, or when your team writes a lot and you don’t want to pay for every request. Gemma 4 and Bielik are free, including for commercial use, and a test on your own computer costs nothing. Buy hardware only when the model passes the test on your tasks.

If you want me to choose the hardware, install the model and set up log reading, see the Local AI model for your business service or write to me. I reply within 24 hours.