Gemma 4 12B release notes / what's new

Google DeepMind has introduced Gemma 4 12B, a mid-sized multimodal model designed to bring agentic intelligence to local hardware. By utilizing a unified, encoder-free architecture, the model provides performance nearing its larger 26B Mixture of Experts (MoE) counterpart while maintaining a memory footprint small enough to run on consumer laptops with 16GB of VRAM or unified memory.

Unified Encoder-Free Architecture

Gemma 4 12B eliminates traditional multimodal encoders to reduce latency and memory usage. Instead of using separate encoders to translate images and audio before passing them to a language model, the vision and audio inputs flow directly into the LLM backbone.

Native Vision Processing

The model replaces the standard vision encoder with a lightweight embedding module. This module consists of a single matrix multiplication, positional embedding, and normalizations, allowing the LLM backbone to handle visual processing directly.

Native Audio Processing

Audio processing is further simplified by removing the audio encoder entirely. The raw audio signal is projected directly into the same dimensional space as text tokens, enabling the model to transcribe, format, and translate voice inputs offline.

Performance and Agentic Capabilities

Gemma 4 12B is positioned as a bridge between the edge-friendly E4B and the 26B MoE model. It delivers benchmark performance that nears the 26B model, which unlocks the ability to handle multi-step reasoning and complex agentic workflows locally.

To support these capabilities, Google is releasing an official Skills Repository, a library of skills specifically designed to enable agents to build with Gemma models.

Deployment and Accessibility

Released under an Apache 2.0 license, Gemma 4 12B is accessible across the developer ecosystem through several channels:

• Local Inference: Supported via LM Studio, Ollama, the Google AI Edge Gallery App, the Google AI Edge Eloquent app, and the LiteRT-LM CLI.
• Weights and Frameworks: Pre-trained and instruction-tuned checkpoints are available on Hugging Face and Kaggle. Implementation is supported through Hugging Face Transformers, llama.cpp, MLX, SGLang, and vLLM.
• Fine-tuning: Efficient fine-tuning is available via Unsloth.
• Cloud Deployment: Production endpoints can be deployed via Google Cloud, including the Gemini Enterprise Agent Platform Model Garden, Cloud Run, and GKE.

Latency Optimization

To reduce inference latency, Gemma 4 12B is equipped with Multi-Token Prediction (MTP) drafters, making it "drafter-ready" for faster response generation.