- Gemma 4 at a glance: what to remember
- Why this is a game-changer for small and medium businesses
- Multimodal: text, image, audio — all built in
- Concrete business use cases
- How to get started without getting lost
- Mistakes to avoid
Gemma 4: an AI that finally runs on your devices, not in the cloud
Imagine an AI assistant that understands your documents, images, and audio files — and never leaves your computer. No internet connection. No data sent to Google. No monthly bill. That's exactly what Google just delivered with Gemma 4, released on April 2, 2026.
For small and medium businesses in the Basque Country, this is a real turning point. We're no longer talking about technology reserved for CAC 40 giants. We're talking about an open-source model you can install on a laptop, a touchscreen kiosk or a local server — one that rivals systems 20 times its size.
Table of Contents
Gemma 4 at a glance: what to remember
Gemma 4 is Google DeepMind's family of open-source models. Four sizes, one philosophy: put artificial intelligence where it's useful — close to you.
The 4 models:
- E2B (2.3 billion parameters) — for modest devices: smartphones, tablets, kiosks
- E4B (4.5 billion) — the smart trade-off for a laptop or desktop PC
- 26B MoE (25 billion, but only 3.8 active per computation) — performance at a lower cost
- 31B (31 billion) — the full model for workstations
What's new? All these models run without an internet connection. Your data stays with you. And the Apache 2.0 license means you can use them commercially, modify them, and integrate them into your tools — without negotiating with Google.
« Gemma 4 is the smartest model, byte for byte, that we've ever created. »
— Google DeepMind, official announcement
The benchmarks confirm it: on the AIME 2026 math test, the 31B model jumps from 20.8% (Gemma 3) to 89.2%. On code resolution, it reaches 80% on LiveCodeBench — a score that puts it on par with much larger models. And in tool use, it jumps from 6.6% to 86.4%. This isn't a marginal improvement. It's a qualitative leap.
Why this is a game-changer for small and medium businesses
Until now, using AI meant two things: sending your data to someone else, and paying a subscription. For a bakery that wants to automate orders, an accounting firm that wants to analyze invoices, or a tradesperson who wants to manage quotes by email — it was either too expensive or too risky.
Gemma 4 changes the game on three fronts:
1. Zero cloud infrastructure cost. No OpenAI subscription, no AWS bill. The model runs on hardware you already own. A MacBook Air is enough for the E4B model. A gaming PC with an NVIDIA graphics card handles the 26B MoE without a problem.
2. Total privacy. Your invoices, quotes and customer data never leave your network. For regulated professions (accountants, lawyers, healthcare), that's a huge strategic advantage.
3. Real customization. Because it's open-source, you can fine-tune the model on your own data. A baker who wants an assistant that understands their recipes can do it — without depending on an outside provider.
And contrary to what people assume, “local” doesn't mean “simplistic.” Gemma 4 understands French (among 140+ languages), processes images and audio, and can even carry out complex multi-step tasks — what's known as agentic workflows.
Multimodal: text, image, audio — all built in
Where previous models only handled text, Gemma 4 is natively multimodal. In practice, that means:
Text + Image: Show it a photo of your stock shelf, and it tells you what's missing. Give it a scanned invoice, and it extracts the amount, the supplier, the date. Show it a floor plan, and it explains the relevant standards.
Audio (on the E2B and E4B models): Talk to it, and it understands. Give it a meeting recording, and it produces a summary. The audio model supports up to 30 seconds of input and covers the main languages.
Video: Process video as a sequence of images — useful for analyzing surveillance footage, product demos or tutorials.
For an SME, this opens up use cases that were impossible before: a document-classification system that reads PDFs and files them automatically, a voice assistant to handle customer calls, or a quality-control tool that checks products by image.
And all of this runs on your local server. No need to send everything to the cloud for “AI” to do the work.
Concrete business use cases
Here are realistic scenarios for SMEs in the Basque Country:
The accountant: Import your PDF invoices. Gemma 4 extracts the data (amounts, VAT, dates, suppliers) and files it automatically. No €50/month SaaS, no data in the cloud. It all runs on a server in your back office.
The tradesperson: You receive a client email with a photo of the job site. Gemma 4 analyzes the image, identifies the problem, and drafts a pre-filled quote for you. All you have to do is approve it.
The retailer: Sort your products by category, generate product descriptions, analyze customer reviews. The model understands business context and can generate tailored responses.
The restaurant owner: A voice assistant handles reservation calls, understands special requests (allergies, occasions), and updates the schedule. Customer data never leaves the premises.
These cases aren't theoretical. They rely on capabilities already tested and documented by the open-source community around Gemma.
How to get started without getting lost
You don't need to be a developer to try Gemma 4. Here are three paths, from simplest to most complete:
Level 1 — Curious (5 minutes)
- Go to Hugging Face and search for Gemma 4
- Choose the E2B or E4B model based on your hardware
- Install Ollama (1-click, Mac/Windows/Linux)
- Run
ollama run gemma4:e4bin your terminal - Ask a question. That's it.
Level 2 — Pragmatic (1 hour)
- Install llama.cpp for optimal performance
- Download the quantized version of the 26B MoE model
- Write a small Python script that processes your documents
- Test it on your real use cases (invoices, emails, quotes)
Level 3 — Ambitious (1 day)
- Set up an inference server (vLLM or llama.cpp server)
- Connect it to your existing tools via the API
- Add tool calling (function calling) to automate tasks
- Deploy on a local server the whole team can use
The key is to start small. Test on a real case, measure the gain, and scale up gradually.
Mistakes to avoid
Mistake #1: thinking local means limited. Gemma 4 rivals models much larger than it. The 31B reaches 84.3% on GPQA Diamond — an advanced reasoning test. Don't underestimate what a local model can do.
Mistake #2: trying to automate everything at once. Start with ONE use case. Master it. Then expand. The best results come from a gradual approach.
Mistake #3: ignoring the hardware. A model runs, but not at the same speed everywhere. On a MacBook Air with E4B, you'll get ~30 tokens/second — comfortable. On a gaming PC with an RTX 4070, the 26B MoE runs at ~50 tokens/second. Match your ambitions to your hardware.
Mistake #4: forgetting security. Even locally, configure access properly. Limit who can query the model. Protect your training data if you fine-tune. Local AI isn't automatically safe — it's simply under your control.
Gemma 4 isn't a marketing promise. It's a tool that works, available today, free, and that genuinely changes the game for businesses that want AI without depending on anyone. The question is no longer “can I afford it?” — it's “what am I waiting for?”
Want to try Gemma 4 on your own use case? Let's talk — we'll guide you step by step.
FAQ — Your questions, our answers
Can Gemma 4 really run on a regular desktop computer?
Yes. The E4B model (4.5 billion parameters) runs on any PC with 8 GB of RAM. The 26B MoE requires a graphics card with 16 GB of VRAM — an RTX 4070 or higher.
Do you need to be a developer to use Gemma 4?
Not necessarily. With Ollama, you can launch the model with a single command. For more advanced use cases (document classification, voice assistant), technical help is recommended — which is exactly what we offer at Mister Anderson.
Does the data really stay local?
Yes. The model runs on your machine. No data is sent to Google or to a remote server. This is one of the main advantages of local AI for regulated professions.
Is Gemma 4 compatible with other tools?
Yes. It works with llama.cpp, Ollama, vLLM, transformers (Python), and is integrated into Hugging Face. The Apache 2.0 license allows any commercial use.
What's the difference between Gemma 4 and paid models like ChatGPT?
ChatGPT remains more powerful for very complex tasks (long-form reasoning, integrated web search). But for 80% of SME use cases (classification, data extraction, writing, image analysis), Gemma 4 is enough — and it's free, and local.