Google Unveils Two Innovative Open LLMs


  • Google unveils Gemma, a series of open-weight models inspired by Gemini.
  • Gemma 2B and Gemma 7B cater to commercial and research needs.
  • Performance benchmarks against competitors like Meta and Mistral are pending.
  • Dense decoder-only architecture promises versatility in application.
  • Developers gain access to pre-packaged Colab and Kaggle notebooks for streamlined usage.
  • Gemma models, while open, are not fully open-source, allowing customization and fine-tuning.
  • Enhanced generation quality enables AI application development on local setups or Cloud environments.
  • Google also releases a responsible generative AI toolkit and a debugging tool for support.

Main AI News:

Just a mere week following the debut of its Gemini series’ latest rendition, Google has proudly introduced Gemma, an all-new collection of streamlined open-weight models. Kicking off with Gemma 2B and Gemma 7B, these fresh additions draw inspiration from the acclaimed Gemini lineup and are poised to serve both commercial and research purposes.

Despite Google refraining from furnishing us with an exhaustive research paper outlining the comparative performance of these models against counterparts from Meta and Mistral, they have assured us of their cutting-edge quality. Described as dense decoder-only models, akin to the architecture underpinning the Gemini and preceding PaLM models, the tech giant promises to unveil comprehensive benchmarks later today via Hugging Face’s leaderboard.

Embarking on your Gemma journey is made seamless, with developers gaining access to pre-packaged Colab and Kaggle notebooks, alongside integrations with Hugging Face, MaxText, and Nvidia’s NeMo. Once primed with pre-training and fine-tuning, these models promise ubiquitous functionality.

While Google emphasizes the openness of these models, it’s crucial to note they aren’t entirely open-source. Janine Banks, speaking ahead of the announcement, underscored Google’s dedication to open-source principles while delineating the distinction in terminology for Gemma models. She remarked, “[Open models] has become pretty pervasive now in the industry,” emphasizing the wide access for customization and fine-tuning, albeit with variations in terms of redistribution and ownership.

Developers can leverage these models for inferencing and fine-tuning as desired, with Google’s team asserting their suitability for a myriad of use cases. Tris Warkentin, Google DeepMind’s product management director, lauded the enhanced generation quality of these models, unlocking novel avenues for AI application development, even on local developer setups or within Google Cloud environments.

The competitiveness of Gemma models vis-a-vis counterparts from Google’s rivals remains to be seen in practical scenarios. In tandem with the model launch, Google is rolling out a new responsible generative AI toolkit, aimed at facilitating the creation of safer AI applications leveraging Gemma models, coupled with a debugging tool for enhanced development support.


The introduction of Gemma by Google signifies a strategic move to expand its offerings in the AI development landscape. With a focus on open-weight models, Google aims to empower developers while maintaining control over model distribution and ownership. The enhanced generation quality and accompanying toolkits indicate Google’s commitment to fostering safer and more efficient AI applications, positioning itself competitively in the market.