On April 2, Google DeepMind released Gemma 4.
This is not the first Gemma, but two things are different from before: it has switched to the Apache 2.0 license, and the 31B version has reached third place on the Arena AI open-source leaderboard.
Four models, how to choose?
| Model | Type | Active Parameters | Context | Special Capability |
|---|---|---|---|---|
| E2B | Dense (on-device optimized) | 2B | 128K | Audio input |
| E4B | Dense (on-device optimized) | 4B | 128K | Audio input |
| 26B A4B | MoE | ~4B | 256K | — |
| 31B | Dense (flagship) | 31B | 256K | — |
E2B and E4B are designed for phones and edge devices, with native audio input (speech recognition) support and multimodal coverage of text, images, and audio. The 26B MoE design works like DeepSeek: 26B total parameters, but only about 4B are activated during inference, offering high cost-effectiveness.
The 31B is the flagship, ranking third among open-source models on the Arena AI text leaderboard, while the 26B MoE ranks sixth. Both large versions outperform competitors with 20 times fewer parameters.
What the Apache 2.0 switch means
Previous Gemma series used a custom license with usage restrictions, requiring separate negotiations with Google for commercial scenarios. After switching to Apache 2.0:
- Free for commercial use, no application needed
- Can modify code and redistribute
- No additional terms required
This aligns with Qwen 3's Apache 2.0 open-source strategy. In the open-source large model licensing landscape, Apache 2.0 is now the standard answer — anyone using restrictive licenses is at a disadvantage.
How practical are the on-device capabilities?
E2B and E4B are designed for phones and Raspberry Pi. According to Google, E4B is 4x faster and uses 60% less power than previous versions. E2B offers 3x faster performance with lower latency.
Given the storage and computing power of current flagship phones, E4B should run smoothly. Google explicitly stated that Gemma 4 will serve as the base model for Gemini Nano 4 on Android devices later this year. In other words, this isn't just for developers — it will eventually reach ordinary users' phones.
How does it compare to Meta Llama 4?
Meta is also pushing Llama 4 at the same time. Both companies are competing in open source, both use MoE or efficient on-device architectures, both support multimodality, and both use the Apache 2.0 license.
Gemma 4's differentiator is on-device: the E series' audio input support and ultra-low power design are areas Llama 4 hasn't focused on yet. If you're a developer working on mobile or IoT scenarios, Gemma 4's E series is worth serious consideration.
Support for over 140 languages is also significant for non-English scenarios. How well Chinese performs will be known once the community runs tests.
Sources: CocoLoop, Gemma 4: Byte for byte, the most capable open models (Google DeepMind Blog); Google Releases Gemma 4 in Four Model Sizes Under Apache 2.0 License (gHacks Tech News); Announcing Gemma 4 in the AICore Developer Preview (Android Developers Blog)