Gemma4 (avenchat.com)

Google has unveiled Gemma 4, a new family of multimodal AI models offering four variants: 31B, 26B A4B, E4B, and E2B. This diverse lineup allows users to balance quality, latency, and hardware costs, catering to various deployment contexts. Notably, the E2B and E4B models feature a robust 128K context, while the larger 31B and 26B A4B variants support an impressive 256K context, making Gemma 4 well-suited for long-document analysis and agent workflows. Furthermore, all models support image input, with the compact E2B and E4B versions also allowing for native audio input, enhancing their utility for edge applications. Gemma 4's significance lies in its open-weight architecture and flexible deployment options, which potentially streamline integration into existing tech stacks. With an Apache 2.0 license, it encourages self-hosting and customization, appealing to teams wary of vendor lock-in. The official memory guidance of 3.2 GB for the E2B model up to 17.4 GB for the 31B model simplifies hardware planning. The array of options, combined with user-friendly entry points like LM Studio and Ollama, positions Gemma 4 as a versatile choice for developers and organizations navigating the evolving landscape of AI and machine learning deployment.