Google's Gemma 4 open-weight family arrives on Amazon Bedrock under Apache 2.0

Google DeepMind's Gemma 4 open-weight model family is now generally available on Amazon Bedrock under the permissive Apache 2.0 license, one of the headline launches at the AWS Summit in New York. The release spans three instruction-tuned variants — Gemma 4 31B, a 26B-A4B Mixture-of-Experts configuration, and a compact E2B — each tuned for strong intelligence-per-parameter so customers can match model size to workload and budget.

The distribution choice is notable: Google placing its open models natively on Amazon's cloud underscores a market shift away from exclusive partnerships toward a multi-model marketplace where customers pick frontier and open weights by fit rather than vendor loyalty. The same week saw OpenAI's GPT-5.5 and Codex reach GA on Bedrock and NVIDIA's quantized DiffusionGemma land on Hugging Face, reinforcing that the major clouds are racing to host the broadest model catalog with native security and governance controls.

The MoE 26B-A4B variant — with only a fraction of parameters active per token — targets the efficiency-per-dollar concern dominating developer conversation this week, as teams react to runaway inference costs on closed frontier models. Apache 2.0 licensing makes Gemma 4 attractive for commercial fine-tuning and on-prem deployment, positioning it against Mistral, Qwen, and DeepSeek in the open-weight tier. Watch for independent benchmarks comparing Gemma 4 against those rivals and how aggressively enterprises adopt it as a cheaper substitute for restricted closed models.