Gemma 4 at a glance
Gemma 4 comes in four sizes, from phone-ready E2B to a 31B dense model, with function calling, up to 256K context and a commercially permissive license.
What changed with Gemma 4
Gemma 4 includes Effective 2B, Effective 4B, a 26B mixture-of-experts model that activates 3.8B parameters per step, and a 31B dense model. All handle images and video, and the two small models also take audio. Context reaches 128K tokens on the edge models and up to 256K on the larger ones, and training covered more than 140 languages.
The models use the Apache 2.0 license. Weights are on Hugging Face, Kaggle and Ollama, with day-one support in tools such as llama.cpp, vLLM, MLX and LM Studio. Google says the unquantized large models fit on a single 80GB H100, and quantized versions run on consumer GPUs.
Gemma 4 in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- Context windowas of · Source for Context window: blog.google (opens in a new tab)
- Up to 256K tokens
How it compares
Other models in the same category that we have covered, using each model's latest recorded value. (Models)
Where to compare Gemma 4
Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.
Leaderboards
General leaderboards
Why Gemma 4 matters for businesses
For clients with strict data rules, Gemma 4 makes it realistic to run capable agents on their own hardware, with function calling and structured JSON output for automation.
Sources for this story
- Gemma 4: Byte for byte, the most capable open models (opens in a new tab)Official
Google · blog.google
- Gemini API changelog (opens in a new tab)Official
Google · aistudio.google.com
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #open-weights
- #gemma
- #on-device
- #llm
