Google just dropped the Gemma 4 technical report, and it's all about stripping away complexity to make large language models run faster and use less memory on phones and edge devices.
Traditionally, multimodal models needed separate encoders to translate images and audio into features the model could understand. Gemma 4’s 12B model takes a radical approach: it eliminates the visual encoder entirely (a 93% reduction) and removes the audio encoder altogether. Raw image patches and audio frames are projected directly into the model’s latent space, solving the memory waste caused by those external translators.
Gemma 4 also streamlines long-context processing and acceleration. By reusing key-value attention (Value=Key), it reduces the KV cache footprint by 37.5% — effectively needing to remember one less feature. For on-device text prediction, the lightweight E2B and E4B variants shrink the searchable vocabulary from 260,000 tokens down to just 4,096, dramatically speeding up decoding.
But there’s a catch: only the 12B model gets the full encoder-free treatment. Other sizes still rely on external encoders for vision and audio.