Google launches Gemma 4 12B, a multimodal AI model that runs on laptops with 16 GB RAM

Key Points

• Google releases Gemma 4 12B, a 12‑billion‑parameter LLM that runs on laptops with 16 GB RAM.
• Model offers near‑parity with the 26‑billion‑parameter Gemma variant in reasoning tasks.
• Built‑in Multi‑Token Prediction (MTP) drafters improve speed and efficiency.
• Vision processing uses a single‑matrix multiplication embedding, removing bulky encoders.
• Audio inputs are projected directly into the text token space, bypassing separate encoding.
• Weights (~18 GB) downloadable from Kaggle and Hugging Face; accessible via LM Studio and Google AI Edge Gallery.
• Gemma 4 12B supports text, image and audio inputs in a native multimodal framework.