Google launches Gemma 2, its subsequent technology of open fashions

Constructed for builders and researchers

Gemma 2 isn’t solely extra highly effective, it is designed to extra simply combine into your workflows:

• Open and accessible: Similar to the unique Gemma fashions, Gemma 2 is on the market beneath our commercially-friendly Gemma license, giving builders and researchers the power to share and commercialize their improvements.
• Broad framework compatibility: Simply use Gemma 2 together with your most popular instruments and workflows due to its compatibility with main AI frameworks like Hugging Face Transformers, and JAX, PyTorch and TensorFlow by way of native Keras 3.0, vLLM, Gemma.cpp, Llama.cpp and Ollama. As well as, Gemma is optimized with NVIDIA TensorRT-LLM to run on NVIDIA-accelerated infrastructure or as an NVIDIA NIM inference microservice, with optimization for NVIDIA’s NeMo to return. You may fine-tune right this moment with Keras and Hugging Face. We’re actively working to allow extra parameter-efficient fine-tuning choices.
• Easy deployment: Beginning subsequent month, Google Cloud prospects will be capable to simply deploy and handle Gemma 2 on Vertex AI.

Discover the brand new Gemma Cookbook, a group of sensible examples and recipes to information you thru constructing your personal functions and fine-tuning Gemma 2 fashions for particular duties. Uncover simply use Gemma together with your tooling of alternative, together with for frequent duties like retrieval-augmented technology.

Accountable AI growth

We’re dedicated to offering builders and researchers with the assets they should construct and deploy AI responsibly, together with by means of our Responsible Generative AI Toolkit. The not too long ago open-sourced LLM Comparator helps builders and researchers with in-depth analysis of language fashions. Beginning right this moment, you should use the companion Python library to run comparative evaluations together with your mannequin and knowledge, and visualize the leads to the app. Moreover, we’re actively engaged on open sourcing our textual content watermarking expertise, SynthID, for Gemma fashions.

When coaching Gemma 2, we adopted our strong inner security processes, filtering pre-training knowledge and performing rigorous testing and analysis in opposition to a complete set of metrics to establish and mitigate potential biases and dangers. We publish our outcomes on a big set of public benchmarks associated to security and representational harms.