Google just dropped Gemma 4 12B, packing real agentic multimodal power into a model small enough for everyday laptops. It sits between the lightweight E4B and the bigger 26B MoE version while adding native audio inputs for the first time in this size range.
Gemma 4 models have now passed 150 million downloads. Developers have used them for everything from wearable robotic arms to enterprise AI security tools.
What makes this version special?
- Unified architecture No separate multimodal encoders. Vision and audio feed straight into the LLM backbone, cutting latency and memory use.
- Strong reasoning It nearly matches the 26B model on benchmarks, making multi-step agentic workflows possible.
- Laptop friendly Runs locally with just 16GB VRAM or unified memory.
- Fully open Apache 2.0 license, plus Multi-Token Prediction drafters for lower latency.
- Run Powerful Agents Locally You get near-top performance at half the memory footprint. Perfect for offline multimodal tasks on consumer hardware.
How It Processes Inputs
Vision uses a super-light embedding module. Audio skips encoders entirely and projects raw signals directly into the token space.
Get Started Now
Test it in LM Studio, Ollama, or Google AI Edge apps. Grab weights from Hugging Face and Kaggle. Build agents faster with the new Gemma Skills Repository. Deploy on Google Cloud when ready.
