Google's Gemma 4 Brings Multimodal Open Models to Everyone

Google has released Gemma 4, a new family of open-source AI models available under the Apache 2.0 licence. The release covers several variants — from compact edge-optimised models to large dense and mixture-of-experts configurations — each capable of handling text, images, and audio natively.

The headline model is a 27B-parameter dense configuration that Google claims ranks third among all open models on major benchmarks. Alongside it sits a 26B-parameter mixture-of-experts variant and two smaller models (E2B and E4B) built for deployment where compute is constrained — on-device applications, edge hardware, anything that cannot route inference through a data centre.

The Apache 2.0 licence is significant. Permissive open licences are increasingly where enterprise adoption actually happens: companies can fine-tune these models on proprietary data and run them without sending traffic through a third-party API. That matters both commercially and for organisations with strict data governance requirements.

Google has billed Gemma 4 as “byte for byte, the most capable open models.” The third-place benchmark ranking gives that claim at least some factual grounding, though benchmark positions shift quickly in this field. What changes more slowly is the practical reality of which models developers actually reach for when they need something they can run, modify, and ship without licensing friction. Gemma 4 is clearly aimed at that space.

The multimodal capability — native handling of text, images, and audio in a single model — is the more interesting long-term development. Most open-source models have historically required separate specialised models for each modality. Integrating them is harder to do well but considerably more useful in production.