Gemma 4 has been released in Google AI Studio.

Google DeepMind has launched Gemma 2, the successor to its popular open-weight Gemma models, with the first release being the 27-billion parameter variant. This model is now available for free in Google AI Studio, providing developers with a streamlined, no-cost platform for experimentation. A key technical upgrade is the expanded 128,000-token context window, enabling the model to process significantly longer documents and conversations. Initial performance evaluations are impressive, with Gemma 2 27B reportedly surpassing Meta's much larger 70-billion parameter Llama 3 model on several benchmarks, including reasoning and coding tasks like HumanEval and GSM8K.

Beyond raw performance, Gemma 2 is engineered for practical efficiency. It utilizes a novel architecture designed for optimal performance on standard GPU hardware, lowering the barrier for deployment. The release in AI Studio includes tools for supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), allowing developers to customize the model for specific use cases. Google has also announced that 9-billion parameter and instruct-tuned variants are coming soon, alongside versions optimized for NVIDIA and Google Cloud TPU hardware, signaling a comprehensive ecosystem rollout.

The launch positions Gemma 2 as a formidable contender in the competitive open-model space, challenging the dominance of models from Meta and Mistral. By offering state-of-the-art performance in a more efficient package and lowering the access barrier through AI Studio, Google is empowering a broader range of builders to innovate with advanced AI without the infrastructure overhead of massive models.