For Norwegian businesses considering AI, the choice often stands between two different paths: cloud-based APIs from OpenAI or Anthropic, or open models that can be run locally. Google's Gemma family, with Gemma 4 launched in April 2026, represents the most mature alternative for the second path.
What is Gemma?
Gemma is a family of open language models from Google, launched in February 2024 and now in its fourth generation. Gemma 4 comes in these sizes (Google's model overview, read 20 November 2026):
| Model | Type | Context window | Built for |
|---|---|---|---|
| Gemma 4 E2B | Small, about 2 billion effective parameters | 128K tokens | Phones, browsers and other small devices |
| Gemma 4 E4B | Small, about 4 billion effective parameters | 128K tokens | Laptops and simple internal tools |
| Gemma 4 12B | Medium, text, image and audio in one model | 256K tokens | A single workstation with a GPU |
| Gemma 4 26B A4B | Mixture of Experts, 26 billion in total | 256K tokens | Servers handling many requests at once |
| Gemma 4 31B | Dense model, 31 billion parameters | 256K tokens | Server GPUs, demanding tasks |
As of May 2026, the Gemma family has over 400 million downloads, according to Google.
Why open models for Norwegian businesses?
1. GDPR and data sovereignty
When you use a cloud-based API, data is sent to the vendor's servers. For Norwegian businesses processing personal data or sensitive business information, this creates legal and practical challenges:
- The Schrems II ruling from the EU Court of Justice in 2020 limits transfer of personal data to the USA
- GDPR Article 44 requires that data leaving the EEA has "adequate protection level"
- Data processing agreements with American companies are becoming increasingly complex
With Gemma run locally or on EU-based infrastructure, all data remains in your control. For a large document archive, document AI for large archives compares this with cloud deployment options.
2. Predictable costs
Cloud APIs bill per token, so the bill grows with usage. Running Gemma yourself costs roughly the same every month however much it is used: the server, the power and the people who run it.
| Cost driver | Cloud API | Self-hosted Gemma |
|---|---|---|
| What you pay for | Tokens in and out | Server or GPU rental, operations and updates |
| When usage doubles | The bill doubles | Same bill until the server is full |
| Biggest risk | Unexpectedly high usage | Skills and time for operations |
Run the numbers on your own volume before you choose. At low usage an API is almost always cheapest; self-hosting pays off when volume is high and steady.
3. No vendor lock-in
With an open model, you own the weights. You can:
- Fine-tune the model on your own data
- Run it on any hardware
- Switch hosting provider without changing code
- Further develop the model internally
Apache 2.0: What does it mean in practice?
Gemma 4 is licensed under Apache 2.0, one of the most permissive open source licenses (Google, April 2026). Earlier generations (Gemma 1 to 3) have their own terms of use from Google (Gemma Terms of Use), so check the license of the model you actually use. For Gemma 4, Apache 2.0 means:
- Commercial use allowed: You can use Gemma in products you sell
- No copyleft: You don't need to open your own source code
- Patent license: The license grants rights to the contributors' patents covering the code
- Sublicensing: You can build on Gemma and sell the result
Technical requirements for local deployment
Hardware
Memory needs depend on model size, quantization and how much text you send in. Google lists the memory requirement per model and quantization in the Gemma 4 model overview; use that table when you size the server. Roughly:
| Model | Typical hardware |
|---|---|
| E2B and E4B | Laptop or phone, can run without a dedicated GPU |
| 12B | Workstation with one GPU |
| 26B A4B and 31B | Server GPU, or a quantized version on a powerful graphics card |
Remember that a long context window uses a lot of extra memory when it is actually filled.
Software
To run Gemma locally, you need:
- llama.cpp or Ollama for simple execution
- Hugging Face Transformers for Python integration
- vLLM or TGI for production serving
Quick start with Ollama
# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Download Gemma 4 E4B
ollama pull gemma4:e4b
# Run interactively
ollama run gemma4:e4bHow to get started?
Step 1: Define the use case
Not all tasks require a 31B model. Start by mapping:
- What kind of text should the model generate? (Email, reports, code, customer service?)
- How important is latency? (Should the user wait 1 second or 10?)
- How important is accuracy? (Creative text vs. medical information)
Step 2: Choose model size
For most Norwegian SMBs, E4B or 12B is a good starting point:
- E4B: Simple text generation and internal tools
- 12B: Customer service, documents with images, more demanding text analysis
- 26B A4B or 31B: Code generation, complex analysis, fine-tuning
Step 3: Set up infrastructure
Alternative A: Own server
- Rent a GPU server from a provider with a data center in the EU (check the provider's current price)
- Own data center: High initial cost, low operating costs
Alternative B: Cloud with EU data center
- Google Cloud europe-west3 (Frankfurt)
- Azure West Europe (Amsterdam)
- AWS eu-north-1 (Stockholm)
Step 4: Test and fine-tune
Download the model, test it on representative tasks, and fine-tune if necessary:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-E4B-it")
# Test in Norwegian
prompt = "Write a short email to a customer asking for feedback."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))Common pitfalls
1. Underestimating infrastructure A 31B model requires significantly more resources than E4B. Many businesses start too large and struggle with performance.
2. Lack of fine-tuning Gemma is trained on general data. For specialized tasks (e.g., Norwegian legal text), fine-tuning on own datasets is required.
3. Security gaps Even local models can leak data if accessible from the internet without authentication. Use VPN/firewall.
4. Underestimating operations Local hosting requires expertise in model serving, updates, and monitoring. This is not "set up and forget".
Summary
Gemma represents a mature, open alternative for Norwegian businesses that want AI control. With the Apache 2.0 license on Gemma 4, local deployment and predictable costs, it addresses the most important concerns around cloud-based AI services.
For most SMBs, E4B or 12B is a good starting point. Begin with a clear use case, test on representative data, and scale gradually as needed.
Next step: Want to know if AIKI can help you set up Gemma for your business? Book a no-obligation consultation.
Sources:
- Gemma 3 announcement (Google, March 2025)
- Gemma 4 announcement (Google, April 2026)
- Gemma models on Hugging Face (Google)
- Gemma 4: models, context window and memory requirements (Google, read 20 November 2026)
- Gemma 4 license and earlier Gemma Terms of Use (Google)
- GDPR Article 44 (EU)




