Google Introduces Gemma 4: Open-Weight AI Models with Multimodal and Reasoning Capabilities

Google has unveiled Gemma 4, the latest addition to its open-weight model family, bringing significant advances in multimodal processing, reasoning, and computational efficiency. The new suite spans a range of scales, from 2.3 billion to 31 billion parameters, with both dense and Mixture-of-Experts architectures designed to balance performance with practical deployment needs.

Multimodal Architecture and Processing#

The Gemma 4 models natively integrate vision and audio capabilities across all model sizes. Most notably, the 12B variant introduces a unified, encoder-free architecture that directly processes raw audio and image patches without intermediate conversion steps. This design choice simplifies the pipeline while maintaining quality, alongside enhanced vision and audio encoders that work across the entire model family.

Thinking Mode for Reasoning#

A key innovation in Gemma 4 is the integration of a "thinking mode," which allows the models to generate internal reasoning traces before producing their final responses. This capability enables more transparent and deliberate problem-solving, particularly beneficial for complex tasks requiring multi-step logic.

Performance and Efficiency Gains#

The model suite delivers measurable improvements in inference speed, memory consumption, and compute requirements. Gemma 4 also extends long-context handling capabilities through architectural refinements. According to the technical report, the models achieve strong performance across STEM reasoning, multimodal tasks, and long-context benchmarks—in some human-evaluated settings, matching or exceeding larger frontier open models.

Open Availability#

As an open-weight model family, Gemma 4 follows Google's approach of releasing model weights publicly, enabling researchers and developers to fine-tune, evaluate, and deploy the models independently.