Models / Gemma
Gemma 4 E4B
Strengths
Native multimodal (text, image, video, audio) at edge sizes. Apache 2.0. ~4B effective inference footprint built to preserve RAM and battery on consumer devices.
Weaknesses
4B class can't match 8B+ on hard reasoning or long-form writing. Audio and video are best-effort, not a substitute for purpose-built pipelines.
Gemma 4 E4B is Google's edge-tier release in the April 2026 Gemma 4 family. The "E" denotes effective parameters: architectural tricks keep the inference footprint at ~4B while drawing on more total parameters during training. Translation: better quality than a vanilla 4B for the same memory budget.
The headline change versus Gemma 3 is native multimodality. E4B accepts text, images, video, and audio out of the box, not as an adapter.
When to pick it
- On-device or edge assistants where every megabyte matters.
- Need a small model that natively handles screenshots, photos, or short audio clips.
- Apache 2.0 with no commercial caveats.
When to skip it
- You have GPU room for an 8B+ model: quality scales, and Qwen3-8B will outperform on most text tasks.
- Heavy multimodal needs (long video, complex visual reasoning): the 26B/31B Gemma 4 variants fit better.