gemma4

Gemma 4 is an open-weights model family created by Google and released under a commercially permissive Apache 2.0 license, derived from the same research and technology that powers Gemini 3. The lineup is engineered for advanced reasoning and agentic workflows, with multimodal input handling that combines text and images for richer interaction. Its release received same-day vLLM support across Google TPUs, AMD GPUs, and Intel XPUs, reflecting a deliberate focus on broad hardware accessibility from phones up to high-scale data center accelerators.

The family spans four versatile sizes tailored to different deployment environments: an Effective 2B (E2B) variant, an Effective 4B (E4B) variant, a 26B Mixture-of-Experts (MoE) configuration, and a 31B Dense model. This range allows practitioners to trade off compute footprint against capability, with the MoE option offering efficiency for high-throughput serving and the dense 31B providing a single-process quality target. The combination of permissive licensing, multimodal grounding, and a size ladder suited to diverse hardware makes Gemma 4 a practical choice for developers building reasoning-heavy assistants, vision-augmented tools, and agent pipelines on open infrastructure.