As AI advances and inference costs soar, smaller models are gaining popularity among developers, and Google is getting in on the action.

On Thursday, the company launched Gemma 4, the next generation of its open models. Google calls Gemma 4 “its most capable open model family yet,” representing a “massive leap” from its predecessor and outperforming bigger models in reasoning, code generation, and complex logic.

To back up the claim, Google shared some of Gemma 4’s strengths:

• Advanced reasoning involving multi-step planning and deep logic
• The ability to support agentic workflows and build reliable autonomous agents
• Support for high-quality offline code
• Native processing of video and images
• Longer context windows
• Training on 140 languages

One of the model’s biggest improvements is its performance relative to its size, or, as Google calls it, “intelligence-per-parameter.” Gemma 4 was sized to run and fine-tune on hardware such as Android devices, laptop GPUs, and IoT devices. To accommodate the different needs, Gemma 4 was released in four different sizes:

• Effective 2B and Effective 4B: The models use a 2B- and 4B-parameter footprint, respectively, during inference to preserve RAM and battery life. Created in collaboration with the Google Pixel team and hardware companies such as Qualcomm and MediaTek, these multimodal models run offline with near-zero latency, according to Google.
• 26B Mixture of Experts (MoE): Ranks as the #6 open model in the world on the industry-standard Arena AI text leaderboard, competing with models 20x its size. Its focus is on latency.
• 31B Dense: Ranks as the #3 open model in the world on the industry-standard Arena AI text leaderboard. Its focus is on “maximizing raw quality,” which Google says is a powerful foundation for fine-tuning.

Unlike Google’s flagship Gemini models, the Gemma models are open-source, encouraging developers to build on and collaborate with them, a move that has already yielded positive results. Since the first Gemma generation launch in February 2024, Google says that it has been downloaded over 400 million times. However, now it also features an Apache 2.0 license for added developer flexibility and control. The model weights are available on Hugging Face, Kaggle and Ollama.

Gemma 4 (31B and 26B MoE) is accessible in Google AI Studio, while Gemma 4 (E4B and E2B) is accessible in the Google AI Edge Gallery. Developers can also use it to power Agent Mode in Android Studio and to build production apps on Android using the ML Kit GenAI Prompt API.

Our Deeper View

Google has taken a nuanced stance in the open vs. proprietary model debate by making select models openly available while keeping its flagship offerings, such as the Gemini family, proprietary. This approach has plenty of merit: it fosters innovation and collaboration within the developer community, which in turn benefits Google by driving higher-quality applications across its various ecosystems, such as Android, all without compromising its intellectual property or its competitive position in the AI race.