Google has announced the release of Gemma 4, a significant advancement in open-weight models, with a notable change in licensing. Gemma 4 is now available under the standard Apache 2.0 license, removing usage restrictions that previously hindered its adoption. This move aligns Gemma 4 with other industry models like Qwen and Mistral, fostering a more open ecosystem.
Gemma 4 boasts four models, divided into workstation and edge tiers. The workstation tier features a 31B-parameter dense model and a 26B A4B Mixture-of-Experts model, supporting extensive text and image inputs. The edge tier includes the E2B and E4B models, tailored for compact devices like phones and laptops.
A standout feature of Gemma 4 is its native multimodality, seamlessly integrating vision, audio, and function calling. These capabilities are deeply embedded in the architecture, offering enhanced flexibility and performance. Additionally, Gemma 4’s models are designed to optimize deployment flexibility, with workstation models supporting consumer-grade GPUs and edge models tailored for efficient inference on smaller devices.
With Gemma 4’s impressive benchmarks across various tasks, including mathematical reasoning and vision processing, Google is setting a new standard in the open-weight model landscape. The introduction of Gemma 4 under the Apache 2.0 license marks a significant milestone, unlocking new possibilities for enterprise AI teams in model deployment and fine-tuning.
Source: VentureBeat