Gemma 4 (Google): small, Apache-licensed open models

Contents

1. What it does for a company
2. Strengths and limits
3. Pricing
4. Questions to ask before choosing Gemma 4
5. Sources

Type	Open-weight model family
Pricing	Published
Website	huggingface.co

Gemma 4, announced on 2 April 2026, moved Google’s open models to the Apache 2.0 licence and covers sizes from phones to a single data-centre GPU⁠12Gemma 4: Byte for byte, the most capable open modelsGoogle · 2 Apr 2026Gemma 4 licence (Apache License 2.0)Google AI for Developers.

Best for: Small and on-device models under Apache 2.0, from phones to a single GPU.

Not the right fit for: Workloads that need frontier-scale reasoning from one model.

What it does for a company

The family comes in E2B and E4B sizes for devices, a 26-billion-parameter mixture-of-experts model that activates 3.8 billion per token, a dense 31-billion-parameter model and a 12B published on Hugging Face in May⁠13Gemma 4: Byte for byte, the most capable open modelsGoogle · 2 Apr 2026google/gemma-4-31B-itGoogle (Hugging Face). The 26B and 31B models take text and images with a 256,000-token context; the smaller ones also take audio, with a 128,000-token context⁠3google/gemma-4-31B-itGoogle (Hugging Face).

The hardware bar is low. Google states that the 26B and 31B weights fit on a single 80 GB H100 in bf16, that E2B and E4B run offline on phones and a Raspberry Pi, and that fine-tuning works through Colab, Vertex AI or on a gaming GPU⁠1Gemma 4: Byte for byte, the most capable open modelsGoogle · 2 Apr 2026. Vendor-reported, the 31B ranked third among open models on the Arena text leaderboard⁠1Gemma 4: Byte for byte, the most capable open modelsGoogle · 2 Apr 2026. The training data is described, with a January 2025 cutoff, but not released⁠3google/gemma-4-31B-itGoogle (Hugging Face).

Strengths and limits

Strengths

• Apache 2.0, with no user caps or naming rules.
• Google states the 26B and 31B models fit a single 80 GB H100 in bf16, and that fine-tuning works on a gaming GPU.
• Vendor-reported: the 31B ranks third among open models on the Arena text leaderboard.

Limits

• Smaller than the frontier-sized families, so less capable on the hardest tasks.
• Weights only: the training data is described but not released.

Choose another option if: You need the strongest coding or agent model (DeepSeek, GLM) or a fully open one (OLMo).

Pricing

Weights are free.

Questions to ask before choosing Gemma 4

• Which of our tasks can a 26B or 31B model handle, and which need a frontier-sized model behind the same router?
• Do we need a model that runs on devices or offline?
• Will we fine-tune on our own hardware, or through Vertex AI?
• How will we measure quality on our own tasks before and after fine-tuning?

See how Gemma 4 compares with the other options in the main guide.

Sources

1. Gemma 4: Byte for byte, the most capable open models. Google, 2 Apr 2026. Accessed 9 Nov 2026.
2. Gemma 4 licence (Apache License 2.0). Google AI for Developers. Accessed 9 Nov 2026.
3. google/gemma-4-31B-it. Google (Hugging Face). Accessed 9 Nov 2026.

Further reading

• Buyer's guide · Updated quarterlyThe best open-source AI in 2027, for companies that want to own itAn analysis of the open-weight model families a company can run today and the layer that turns one into its own, compared on seven published criteria, with sourced licences, costs and the established path from rented AI to owned AI.
• Practical guideHow to train your own AI model on company dataMost of the work in training a company model happens before any training starts. A step-by-step account of the data work, the choice of base model, the training methods and the rules that apply in Europe.
• Licence guideOpen source vs open weights: what each licence lets a company doAlmost every model sold as open-source AI publishes its weights but not its training data. The difference rarely matters for internal use. It matters a great deal for companies that resell a model or build a product on it.

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski