Cloud Hosting

Today Google DeepMind unveiled Gemma, its new 2B and 7B open source models built from the same research and technology used to create the companys recently-announced Gemini models.

The Gemma models will be released with pre-trained and instruction-tuned variants, Google DeepMind said in a blog post. The model weights will be released with a permissive commercial license, as well as a new Responsible Generative AI toolkit.

Google is also providing toolchains for inference and supervised fine-tuning (SFT) across all major frameworks: JAX, PyTorch, and TensorFlow through native Keras 3.0. There are ready-to-use Colab and Kaggle notebooks, and Gemma is integrated with Hugging Face, MaxText, and NVIDIA NeMo. Pre-trained and instruction-tuned Gemma models can run on a laptop, workstation, or Google Cloud with deployment on Vertex AI and Google Kubernetes Engine.

Nvidia also announced today that in collaboration with Google it had launched optimizations across all NVIDIA AI platforms, including local RTX AI PCs, to accelerate Gemma performance.

The AI Impact Tour NYC

Well be in New York on February 29 in partnership with Microsoft to discuss how to balance risks and rewards of AI applications. Request an invite to the exclusive event below.

Jeanine Banks, vice president and general manager of developer X and head of developer relations at Google, told VentureBeat at a press briefing that the Gemma models felt like a continuation after Googles history of open sourcing tech for AI development, from tools like TensorFlow and Jax to other models and AI systems like PaLM2 and AlphaFold, leading up to Gemini.

She also said that through feedback during the development of the Gemini models, Google DeepMind gained a key insight, which is, in some cases, developers will use both open models and APIs in a complementary way in their workflow depending on the stage of the workflow that theyre in.

As developers experiment and do early prototyping, she explained, it may be easy to start with an API to test out prompts, then turning to customize and fine-tune with open models. We felt that it would be perfect if Google could be the only provider of both APIs and open models to offer the widest set of capabilities for the community to work with.

Tris Warkentin, director of product management for Google DeepMind, told VentureBeat at the press briefing that the company will be releasing a full set of benchmarks evaluating Gemma against other models, which anyone can see on the OpenLLM leaderboards right away.

We are partnering with both Nvidia and Hugging Face, so pretty much any benchmark that is in the public sphere has been run against these models, he said. It is a fully transparent and community open kind of an approach, so it is something that were actually quite proud of because when you look at the numbers, I think weve done a pretty darn good job.

Warkentin also emphasized Gemmas safety: These all have been extensively evaluated to be the safest models that we could possibly put out into the market at these sizes, along with pre-training and evaluation, he said.

The Google DeepMind blog post said that Gemma is designed with our AI Principles at the forefront. As part of making Gemma pre-trained models safe and reliable, we used automated techniques to filter out certain personal information and other sensitive data from training sets. Additionally, we used extensive fine-tuning and reinforcement learning from human feedback (RLHF) to align our instruction-tuned models with responsible behaviors. To understand and reduce the risk profile for Gemma models, we conducted robust evaluations including manual red-teaming, automated adversarial testing, and assessments of model capabilities for dangerous activities. These evaluations are outlined in our Model Card.*

In addition to safety, Warkentin emphasized the role of the open ecosystem in fostering responsible AI.

We think it is really critical we need diverse perspectives from developers and researchers worldwide, in order to get the right feedback and build even better safety systems, he said. So part of the open model journey is to make sure that were integrating [those perspectives] and that feedback, that communication with the community, is a critical part of the way that we view the value of this project.

VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.

See the original post here:
Google DeepMind jumps back into open source AI race with new model Gemma - VentureBeat