Overview
- Google DeepMind released Gemma 4 12B, which it unveiled on June 3, 2026, as a mid-sized multimodal model that the company says can run on consumer laptops with roughly 16 GB of VRAM or unified memory.
- The model uses an encoder-free architecture that feeds vision and audio inputs directly into the language-model backbone, replacing traditional image and audio encoders with lightweight embedding modules.
- Gemma 4 12B ships with Multi-Token Prediction (MTP) drafters to reduce latency by using idle CPU/GPU cycles to predict future tokens, a feature included out of the box for this 12 billion parameter variant.
- Google released the model weights at about 18 GB under an Apache 2.0 license and made checkpoints and tooling available on Hugging Face, Kaggle, LM Studio and the Google AI Edge Gallery, plus a Skills Repository for building agents.
- Industry observers say the launch follows a broader move toward edge-first agentic AI that can cut per-query cloud costs and keep data local, but independent benchmarks and real-world tests of accuracy, multilingual behavior and agentic limits are still pending.