On April 2, Google DeepMind released Gemma 4, its new family of open models. The models are stronger than Gemma 3, as you would expect. The bigger news is the license: Gemma 4 is released under Apache 2.0, replacing the custom Gemma terms that earlier versions shipped with.
Four Models, From Phones to Workstations
- E2B and E4B — around 2 and 4 billion effective parameters, aimed at phones and small devices, with up to 128K tokens of context and voice input for on-device speech recognition.
- 26B mixture-of-experts — only 3.8B parameters active per token, for low latency on larger hardware.
- 31B dense — the quality-first model, intended as a base for fine-tuning. Unquantized, it fits on a single 80 GB GPU; quantized versions run on consumer graphics cards.
All four take images and video as input, the larger two support 256K tokens of context, and Google says the family covers over 140 languages. Function calling, structured JSON output and system instructions are supported natively — the features you need to use a model inside an application rather than as a chatbot.
Why the License Matters
Open-weight models have often come with licenses that are open in name and complicated in practice: use restrictions, acceptable-use policies that change, terms a legal team has to read before anyone can ship. Apache 2.0 is the opposite — a standard license most companies have already approved for countless dependencies.
For product teams, that removes a step that often decides whether a model gets used at all. A model that is slightly better on a benchmark but needs a legal review loses, in practice, to one that can be added to a project today.
The Small Models Are the Interesting Ones
Google says the E2B and E4B models were optimized with Qualcomm, MediaTek and the Pixel team for phones and boards like the Raspberry Pi, and that they will underpin the next generation of Gemini Nano. That makes them a preview of what on-device AI on Android will look like.
We have written before about why we like on-device inference: privacy, latency, offline use and no per-request cost. Models of this size, under a permissive license, with native audio input and tool calling, cover a growing share of real features — classification, extraction, short summaries, speech commands — without a network round trip.
What to Try
- Run E2B or E4B through Google AI Edge Gallery or Ollama on your own hardware, with your own prompts, before reading another benchmark.
- Look for features in your product that currently call a cloud model for a small, well-defined task. Those are the first candidates to move on-device.
- If a license review blocked an open model before, it may be worth asking again.