Gemma 4 on Amazon SageMaker: The NVIDIA T4 Decodes at 0.8x of the L4 With the Same Answers

Gemma 4 was deployed on Amazon SageMaker using NVIDIA T4 and L4 GPUs. Testing shows the T4 decodes at 0.77x to 0.82x the speed of the L4 while providing identical answers, with the 12B model being the largest compatible build. Comparisons between SageMaker endpoints and EC2 instances using vLLM 0.30.0 reveal that EC2 costs 0.71x per hour. Python MCP tools were developed to simplify the management of these vLLM hosted deployments across us-east-2.