Google just released Gemma 4 — the next generation of their open model family. And this time, they’re not messing around.
Four Models for Every Use Case
The Gemma 4 family consists of four models: two compact variants with 2 and 4 billion parameters for edge devices and smartphones, plus two larger models — a 26-billion-parameter Mixture-of-Experts and a 31-billion-parameter Dense model for more powerful hardware.
All four models can process video and images. The two smaller ones also understand audio and speech — making them especially interesting for on-device applications.
Impressive Benchmark Results
The 31B model claimed third place on Arena AI’s text leaderboard, with the 26B MoE model landing at sixth. These models are competing against — and beating — models 20 times their size. On a per-parameter basis, Gemma 4 is one of the most capable open models available.
Apache 2.0 — Truly Open
Gemma 4 ships under the commercially friendly Apache 2.0 license. No restrictions on commercial use, no fine print. Developers can freely deploy, fine-tune, and integrate these models into their products.
Why This Matters
Gemma 4 forms the foundation for the next generation of Gemini Nano. Code you write for Gemma 4 today will automatically work on Gemini Nano 4 devices coming later this year. Google is building a bridge between their open and proprietary model lines.
For developers building local or on-device AI applications, Gemma 4 is one of the most compelling offerings available right now. The combination of multimodality, performance, and licensing is hard to beat.
Sources: