Google DeepMind describes Gemma 4 as its most intelligent open models, built from Gemini 3 research and technology to maximize intelligence per parameter.
Google DeepMind's Gemma 4 page groups E2B and E4B as compute- and memory-efficient models for mobile and IoT devices, and groups 12B, 26B and 31B as models aimed at personal computers.
The Gemma 4 Technical Report is listed on arXiv as arXiv:2607.02770, submitted on 2 July 2026 and revised on 24 July 2026.
The Hugging Face model page for google/gemma-4-12B lists the model under the google organization, links to the Gemma 4 Technical Report, and lists the license as apache-2.0.
Google's developer blog says Gemma 4 12B is a dense multimodal model with a unified, encoder-free architecture that feeds multimodal data straight into the LLM backbone instead of using separate vision and audio encoders.
The same Google developer post says Gemma 4 12B is the first medium-sized Gemma model capable of natively ingesting audio, and says it is small enough to run locally on dedicated GPU laptops with 16GB VRAM or unified memory.
Google says developers can get pre-trained and instruction-tuned Gemma 4 12B checkpoints from Hugging Face and Kaggle, run local inference with Hugging Face Transformers, llama.cpp, MLX, SGLang and vLLM, or deploy through Google Cloud Model Garden, Cloud Run and GKE.
For local-AI builders, the practical change is a Google open-weight model line tied to consumer-device inference, multimodal input and standard local serving stacks; independent Shadowfetch latency or power measurements for Gemma 4 are DATA UNAVAILABLE in the cited sources.
