Strengths
- The Apache 2.0 license allows commercial use and changes to the model. Earlier Gemma generations came with Google's own terms.
- Quantized, meaning stored with lower numeric precision, it runs on a single GPU. The 4-bit weights take about 17.5 GB of memory, the 8-bit ones 34.9 GB.
- Strong for its size: Google reports 84.3% on GPQA Diamond and 80.0% on the LiveCodeBench v6 coding test.
- It handles more than 35 languages out of the box and saw over 140 during training.
- You can switch thinking on when you need it and leave it off for simple questions.
Weaknesses
- Its knowledge stops in January 2025, so without search it knows nothing about later events.
- The context is 256K tokens, while the Gemini API models handle about a million. This size does not accept audio input.
- It scores only 19.5% on Humanity's Last Exam without tools, which shows the gap to the strongest models.
- Running it yourself takes a GPU with enough memory and someone to set up and maintain the model.
Best for
- running on your own company server
- fine-tuning for your own tasks
- a local assistant on a workstation with a GPU
- data that must not leave the company
Not ideal for
- questions about recent events without search connected
- documents longer than 256K tokens
How to use it in Europe
You can download the weights for free from Hugging Face or Kaggle and run them in apps such as Ollama or LM Studio, or on your own server. The Apache 2.0 license has no geographic limits, so the model can run inside your company in the EU and your data stays in house. In the Gemini API and Google AI Studio, Gemma 4 is only on the free tier, with no paid tier for it. For production use, your own server is the better option.
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| MMLU Pro | 85.2% | ai.google.dev ↗, April 2026 |
| GPQA Diamond | 84.3% | ai.google.dev ↗, April 2026 |
| LiveCodeBench v6 | 80.0% | ai.google.dev ↗, April 2026 |
| AIME 2026 (no tools) | 89.2% | ai.google.dev ↗, April 2026 |
| Codeforces ELO | 2150 | ai.google.dev ↗, April 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- Gemini 4 Carbon: Google's most capable model, announced on 28 October 2026. So far only selected cybersecurity experts have access.
- Gemini 4 Argon: Google's most capable publicly available model, out in October 2026. Built for demanding analytical work where accuracy matters and made-up answers are costly.
- Gemini 3.8 Flash: Google's fast, lower-cost model for everyday work, coding and document analysis. You can try it in the API for free.
- Gemini 3.5 Flash-Lite: A very fast, low-cost Google model for large volumes of simple tasks, such as pulling data out of documents.
- Gemma 4 26B A4B: Google's faster open model for your own server. Close to the larger Gemma 4 31B in quality, but quicker to respond.
Similar models from other vendors
- Muse Glimmer 30B (Meta): Meta's small open model under Apache 2.0 that runs on a single graphics card, even offline.
- Llama 4 Scout (Meta): The smaller Llama 4 from April 2025, very cheap through providers. Its license grants no rights to people or companies in the EU.
- Ministral 3 14B (Mistral AI): Mistral's small open model under Apache 2.0 that runs on a single GPU, even without an internet connection.
- Nemotron 3.5 Bolt 74B-A9B (NVIDIA): NVIDIA's open model for AI agents on your own powerful PC, small enough to fit in 24 GB of graphics card memory.
- GPT-6 Luna (OpenAI): OpenAI's cheapest model, made for high volumes of simple tasks, sorting text and low-cost chatbots.
- gpt-oss-20b (OpenAI): OpenAI's smaller open model that runs on a computer with 16 GB of memory, even without internet access.