Google Unveils Gemini 4 AI Models with Open Licensing

Original: Google announces Gemma 4 open AI models, switches to Apache 2.0 license

Ars Technica•Apr 2, 20262 min

In a Nutshell

Google has launched the Gemini 4 AI models, offering developers improved performance and local usability. Transitioning to an Apache 2.0 license, these models come in four sizes optimized for various hardware, including mobile devices, and aim to enhance processing speed and efficiency.

Key Points

01Gemini 4 models are now available under the Apache 2.0 license, enhancing developer freedom.
02The models are designed for local use, with options for high-end GPUs and mobile devices.
03Performance improvements include reduced latency and optimized memory usage.
04Gemini 31B is expected to rank third among top open AI models.
05The new models are more cost-effective to run compared to previous versions.

In-Article Ad

Full Summary

Google has announced the launch of its Gemini 4 AI models, which significantly enhance local usability and performance for developers. Transitioning from a custom license to the more flexible Apache 2.0 license, Gemini 4 includes four model sizes optimized for various hardware configurations. The two larger variants, 26B Mixture of Experts and 31B Dense, are designed to run on high-end GPUs, specifically the 80GB Nvidia H100, while still being adaptable for consumer GPUs when quantized. Additionally, the smaller models, Effective 2B (E2B) and Effective 4B (E4B), are tailored for mobile devices, ensuring low memory usage and near-zero latency. Google claims that these new models outperform their predecessors, with Gemini 31B expected to secure a place as the third most capable open AI model. Overall, the Gemini 4 models promise to deliver enhanced processing capabilities while being more accessible and cost-effective for developers.

Key Figures

80GB

Required GPU memory for running larger models

20,000

Approximate cost in USD of the Nvidia H100 AI accelerator

3.8 billion

Active parameters used in inference mode for the 26B Mixture of Experts model

2 billion / 4 billion

Effective parameters for mobile models E2B and E4B

In-Article Ad

Why It Matters

The availability of Gemini 4 models under an open license will empower developers to create more efficient AI applications on local hardware, potentially transforming how AI is integrated into everyday devices.

Who is affected

Developers, businesses utilizing AI, and users of mobile and local computing devices.

What to know

Developers should explore the new Gemini 4 models to leverage their capabilities for local AI applications.

FAQ

The Apache 2.0 license allows developers greater freedom to use, modify, and distribute the Gemini 4 models without stringent restrictions.

Gemini 4 models are designed to be more efficient, with reduced latency and optimized for local hardware, making them more capable than the earlier Gemini 3 models.

✦