# Unveiling the Latest Gemma Model Advancements

Kathleen Kenealy, Google DeepMind | AI Engineer World's Fair 2024 | 16:25

Source: https://www.youtube.com/watch?v=Xmkl27AM2VQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/unveiling-the-latest-gemma-model-advancements
Published: 2025-02-09
Tags: benchmarks, fine-tuning, multimodal, open-models

## TL;DR
- Gemma models are open, lightweight models built from the research and technology used for Gemini, with safety work applied from the start of development.
- The Gemma family includes models for general language tasks, code, long-context inference, and image-text tasks, with broad framework and hardware support.
- Gemma 2 adds 9-billion and 27-billion-parameter models, stronger performance, improved documentation, and a focus on easy downstream fine-tuning.

## Summary
Kathleen Kenealy introduces Gemma as Google DeepMind's family of open, lightweight models built from the research behind Gemini. She describes the team's approach to safety, including manual data inspection and safety evaluations throughout model development. The talk reviews the growing Gemma family, including CodeGemma, RecurrentGemma, and PaliGemma, then focuses on Gemma 2. The new 9-billion and 27-billion-parameter models use the same tools and frameworks as earlier releases, while improving performance, documentation, and fine-tuning support. Kenealy says Gemma 2 models outperform larger models on several benchmarks and that the 27B model compares with Llama 3 70B in performance. She also points developers to Google AI Studio, the Gemma Cookbook, documentation, open-model repositories, and research credits. Her motivation is personal: as an academic researcher, she wanted stronger open foundational models for developers and researchers.

## Key ideas
### Gemma is built as an open model family from Gemini research
[02:41](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=161s)
Kenealy describes Gemma as Google DeepMind's family of open-source, lightweight models, built from the same research and technology used for Gemini. She frames the project around giving the open-source community access to capable models that developers can inspect, adapt, and deploy. The models are available across hardware and software environments, including TPUs, GPUs, and local devices. She says Gemma is supported by TensorFlow, JAX, Keras, PyTorch, Ollama, Transformers, and other tools. Its open access and license let developers decide how to use and extend the models.

### Safety work starts before the model is trained
[03:26](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=206s)
Kenealy says the Gemma team treats safety as a design requirement from the first day of development. The team manually inspects datasets for quality and safety, then evaluates models during early experiments and ablations. Those evaluations help the team choose training methods that produce safer models. At the end of development, Gemma models go through the same state-of-the-art safety evaluations used for Gemini models. Kenealy presents this process as a way to give developers confidence that a customized Gemma deployment can remain responsible across different applications.

### The Gemma family covers general language, code, long contexts, and vision
[05:25](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=325s)
The first Gemma releases were foundational language models. CodeGemma followed with fine-tuning for code generation and code evaluation. RecurrentGemma uses a state space architecture aimed at faster and more efficient inference, especially with long contexts, and expanded from a 2B model to a 9B model. PaliGemma adds image-text capabilities by combining a SigLIP vision encoder with a Gemma 1.0 text decoder. Kenealy lists image and video captioning, visual question answering, object detection, and object segmentation among its uses. PaliGemma is available in a 2B size, with pretrained weights and task-specific variants.

### Gemma 2 keeps the earlier tooling while increasing model size and performance
[08:36](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=516s)
Gemma 2 launches in 9B and 27B parameter sizes. Kenealy says both are highly competitive for their size and outperform models that are two to three times larger. Developers who already use Gemma 1 can switch to Gemma 2 with only a few lines of code because the models use the same tools and frameworks. Gemma 2 also keeps broad compatibility with TensorFlow, JAX, Transformers, and other previously supported tools. Kenealy says the team improved documentation with more guides, tutorials, and examples for inference and efficient fine-tuning.

### Fine-tuning was a direct design target for Gemma 2
[10:17](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=617s)
The Gemma team studied how core modeling choices affect downstream fine-tuning. Kenealy says this work was intended to make Gemma 2 easier to customize for individual use cases. The release includes documentation for inference, fine-tuning, and evaluation across frameworks. The 27B model is also available in Google AI Studio, where developers can select Gemma 2 and begin experimenting with prompts. This gives users a browser-based way to try the model before building a more involved application.

### Gemma 2 competes with much larger open models in evaluations
[11:47](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=707s)
Kenealy says the 27B model performs in the same ballpark as Llama 3 70B and outperforms Grok models on several benchmarks, sometimes by a significant margin. She also describes human evaluations because academic benchmarks do not always predict how models will work in users' applications. In those evaluations, she says Gemma models were consistently preferred over other open models, including larger ones. She identifies Gemma 2 27B as the highest-ranked open model of its size at the time of the talk, while noting that an Mistral Large model scored higher on LMSYS.

### The Gemma Cookbook turns model usage into reusable recipes
[13:15](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=795s)
The Gemma Cookbook on GitHub contains 20 recipes that range from easy examples to advanced applications. Kenealy says the repository accepts pull requests, so developers can contribute examples of what they build and help share those approaches with other users. She also points developers to Google Cloud credits for research, expanded documentation, framework-specific guides, downloadable model repositories, Discord, and social channels. These resources cover inference, fine-tuning, and evaluation rather than only prompt experimentation.

### Kenealy's motivation is to improve open access to capable foundational models
[14:22](https://www.youtube.com/watch?v=Xmkl27AM2VQ&t=862s)
Kenealy says she has worked on Gemma for almost two years. She began because, as an academic researcher, she was disappointed that open foundational language models lagged behind the rapid progress of proprietary models. She describes Gemma as a project she would have wanted during her time in the open-source community. Her closing request is for developers and researchers to share projects, research, feedback, and suggestions for improving future releases through GitHub, Hugging Face, social media, and conversations at events.

## Notable quotes
- "Gemma is Google DeepMind's family of open-source lightweight state-of-the-art models which we build from the same research and technology used to create the Gemini models." (02:41)
- "The first is that Gemma models were built to be responsible by Design." (03:26)
- "Gemma 2 comes in both a 9 billion parameter size and a 27 billion parameter size." (08:36)
- "The Gemma Cookbook is available on GitHub now and contains 20 different recipes ranging from easy to very Advanced applications of how to use the Gemma models." (13:15)
- "I started working on this project because I as a researcher in Academia was disappointed to see how far behind open foundational LLMs were compared to the rapid improvements we were seeing in proprietary models." (15:02)

## Tools & references mentioned
- Google DeepMind
- Gemma
- Gemini
- CodeGemma
- RecurrentGemma
- PaliGemma
- SigLIP
- TensorFlow
- JAX
- Keras
- PyTorch
- Ollama
- Transformers
- Google AI Studio
- Google Cloud
- Llama 3 70B
- Grok
- Mistral Large
- LMSYS
- Gemma Cookbook
- Hugging Face

## Who should watch
- You are choosing an open language model and need to compare model size, framework support, and deployment options.
- You want to fine-tune Gemma for a code, vision, chat, or long-context application and need starting points from the model team.
- You work on open-model research and want to understand the team's safety process and available research support.

## Related talks

- [Gemma 4 Deep Dive](https://aietalks.com/talks/gemma-4-deep-dive) (Cassidy Hardin, Google DeepMind, 19:03)
- [Accelerating AI on Edge](https://aietalks.com/talks/accelerating-ai-on-edge) (Chintan Parikh & Weiyi Wang, Google DeepMind, 23:58)
- [Sovereign Escape Velocity: Ownership with Open Models](https://aietalks.com/talks/sovereign-escape-velocity-ownership-with-open-models) (Gus Martins & Ian Ballantyne, Google DeepMind, 20:52)
- [How Google DeepMind Is Researching the Next Frontier of AI for Gemini](https://aietalks.com/talks/how-google-deepmind-is-researching-the-next-frontier-of-ai-for-gemini) (Raia Hadsell, Google DeepMind, 20:37)
- [A year of Gemini progress + what comes next](https://aietalks.com/talks/a-year-of-gemini-progress-what-comes-next) (Logan Kilpatrick, Google DeepMind, 11:58)