Gemma 4: Google's Open-Source LLM — What's New and How It Compares

Emmanuel Ekunsumi · 4 min read · 2026-07-13

Google's Gemma series is its open-source LLM line — models available for anyone to download, fine-tune, and deploy. Gemma 4 continues the series with improved reasoning, longer context, and better instruction following over Gemma 3.

Gemma 4 model sizes

Model	Parameters	Context	VRAM needed	License
Gemma 4-1B	1B	32K	2GB	Gemma Terms
Gemma 4-4B	4B	128K	4GB	Gemma Terms
Gemma 4-12B	12B	128K	10GB	Gemma Terms
Gemma 4-27B	27B	128K	20GB	Gemma Terms

What's new in Gemma 4

Multimodal support — the 12B and 27B variants support image inputs alongside text
Longer context — 128K token context window on all models except 1B
Better instruction following — improved adherence to system prompts and output format constraints
Improved coding — meaningfully higher HumanEval scores vs Gemma 3
Distillation from Gemini — trained with knowledge distillation from Google's frontier Gemini models

Gemma 4 vs the competition

Model	Size	MMLU	HumanEval	Context
Gemma 4-27B	27B	82%	58%	128K
Llama 3.2-8B	8B	73%	51%	128K
Mistral-7B	7B	63%	31%	32K
Qwen2.5-7B	7B	75%	53%	128K
Phi-4 (3.8B)	3.8B	79%	62%	16K

Running Gemma 4 locally

# Via Ollama
ollama pull gemma4        # 4B default
ollama pull gemma4:12b    # 12B with vision
ollama pull gemma4:27b    # 27B full size

ollama run gemma4

Gemma license considerations

Gemma uses Google's custom Gemma Terms of Use — not a standard open-source license. Key restrictions: can't use for training models that compete with Google products, must attribute Gemma in downstream products. For most commercial use cases this isn't a problem, but check the terms before building on Gemma for any sensitive use case.

Tip: Gemma 4 runs on Ollama's OpenAI-compatible local server, so you can wrap it with Tokoscope to track usage alongside cloud API calls.

Track Gemma and cloud LLM costs in one place

Tokoscope works with any OpenAI-compatible endpoint. Free to start.

Get started free →