Google has rejoined the open model map with a forceful entry: the Gemma 4 family, released on April 2, 2026.
The short version, focused on the part we care about: Gemma 4 is a new open model family from Google released under Apache 2.0, and all four models natively process images and video. Google also says the E2B and E4B variants add native audio input, the edge models carry a 128K context window, and the larger models go up to 256K. In their launch writeup they specifically call out visual tasks like OCR and chart understanding, which is the right place for an open multimodal family to start. Read Google’s full launch post.
The models
- Gemma 4 E2B
- Gemma 4 E4B
- Gemma 4 26B-A4B
- Gemma 4 31B
The two small models are the obvious edge story. The 26B-A4B is the MoE latency play. The 31B is the dense quality play.
Google’s launch graphic showing Gemma 4’s open-model chat arena positioning.
Source: Google, Gemma 4 launch post.
All four Gemma 4 models served on Overshoot on the day of their release.
More to come: we will put these models through the same tests we use for the rest of the open model map: deployment behavior, latency, and where the vision capabilities hold up in practice.