Google's Gemma 4: Open, Capable, and Ready for Enterprise

Google DeepMind's Gemma 4 represents a major shift in open-weight model accessibility, now shipping with a standard Apache 2.0 license. This move eliminates previous usage restrictions and legal review hurdles that had steered enterprises toward other models. The Gemma 4 family includes four distinct models, organized into 'workstation' and 'edge' tiers, offering flexibility from cloud deployments to devices.

The workstation tier features a 31B-parameter dense model and a 26B Mixture-of-Experts (MoE) model, both supporting text and image input with extensive 256K-token context windows. The 'edge' tier comprises compact E2B and E4B models designed for phones and laptops, capable of processing text, image, and audio with 128K-token context windows.

Architectural innovations include MoE models with numerous small experts to optimize inference costs and hybrid attention mechanisms for large context windows. Native multimodality is a key differentiator, with integrated vision, audio processing, and function calling capabilities built into the architecture from the ground up, rather than as add-ons.

These models demonstrate significant performance gains across benchmarks, with the 31B dense model achieving strong scores in mathematical reasoning and coding. The MoE model offers comparable reasoning with substantially lower inference costs. The edge models also outperform previous generations, making them highly capable for on-device applications. The permissive Apache 2.0 license ensures these derivatives can be deployed commercially, opening new avenues for enterprise adoption.