Google Introduces Gemma 4 12B, a Local Multimodal AI Model

Google has introduced Gemma 4 12B, a multimodal artificial intelligence model designed to run directly on laptops. It can process text, images, and audio, bringing advanced reasoning and automation tasks to devices with limited resources.

The model sits between the smaller, more efficient Gemma 4 E4B and the 26-billion-parameter model based on a mixture-of-experts architecture. According to Google, Gemma 4 12B comes close to the latter's performance in standard tests while using less than half the memory.

Multimodal AI without separate encoders

The main innovation is its architecture. Multimodal models typically use separate components to convert images and audio into information the language model can understand. These components, known as encoders, add memory usage and processing time.

Gemma 4 12B does away with them almost entirely. Visual and audio information enters the model's core more directly:

• For images, it uses a lightweight module with matrix multiplication, positional information, and normalization.
• For audio, it projects the original signal into the same space as text tokens, the units the model uses to process a sentence.

The result is a model that can combine written instructions with images and sound without relying on a chain of independent systems. It is also the first medium-sized model in the Gemma family with native audio input.

Designed to run on your laptop

Google says Gemma 4 12B can run locally with 16 GB of unified memory or graphics memory, depending on the device. In practice, this would make it possible to use the model without sending every query to a remote server.

That opens the door to tools that analyze documents and images, listen to instructions, carry out several steps on their own, or automate tasks on your computer. For example, a local assistant could receive a screenshot, interpret what is happening, and help you solve the problem without the content leaving your device.

The model also includes Multi-Token Prediction, a technique that anticipates several tokens at once to reduce latency. Put simply, it can respond with less delay in interactive applications.

Available to developers

Gemma 4 12B is released under the Apache 2.0 license, which allows it to be used and adapted with few restrictions. Google offers pretrained and instruction-tuned versions through Hugging Face and Kaggle.

It can also be tested with tools such as:

• LM Studio and Ollama.
• Google AI Edge Gallery and Google AI Edge Eloquent.
• The LiteRT-LM command-line interface.
• Hugging Face Transformers, llama.cpp, MLX, SGLang, and vLLM.

Developers can fine-tune the model with Unsloth and deploy it on Google Cloud, including Cloud Run, GKE, and Model Garden within Gemini Enterprise Agent Platform. Google has also published a skills library so other agents can build applications with Gemma models.

The Gemma family has already surpassed 150 million downloads, according to Google. The arrival of Gemma 4 12B points to a specific trend: models capable of seeing, listening, and carrying out tasks are beginning to fit on a personal computer. The important question now is how much of that performance holds up in real-world use, and which applications can run quickly and reliably without depending on the cloud.