Google releases new family of open models: Gemma

Small language models

Google's new models show that 2024 is the year of small language models (SLMs) as well as large language models, Gartner analyst Chirag Dekate said.

"In a GenAI era, when enterprises need to be able to create value not just from LLMs, but also from SLMs that can be contextualized in their data context, at a lower price point, which matters to them, that becomes incredibly important," Dekate said.

Smaller language models enable enterprises to get higher levels of precision, said R. "Ray" Wang, founder of Constellation Research.

"It's going to be in these smaller language models where people are actually going to start putting stuff together," Wang said. "The large language models are going to capture all the stuff that's publicly available, but there's going to be data in private networks that's going to require smaller language models."

Since Gemma is a small model, it could become more popular than Meta's open source model family, Llama, AI analyst Mark Beccue said.

"They've published a 2 billion-parameter model -- that's the smallest one I've seen," Beccue said.

Despite its competitive relationship with Meta, Google still offers Llama 2 on Google Cloud. That shows how Google is committed to offering generative AI model options to enterprises, Dekate added.

"What Google is trying to communicate here is you get access to not just open models, but lower-cost models, and more importantly, access to better innovation faster through a Google Cloud ecosystem," he said.

Iris.AI and Gemma

Iris.AI is an AI startup interested in Google Gemma. Iris provides a platform and creates models that use generative AI technology to analyze scientific documents for researchers.

Iris' current generative model is based on the Llama 13 billion-parameter architecture and pretrained on scientific documents.

"What we find very nice about the Llama 2 models and Llama architecture is that it's quite easy to operate," co-founder and CTO Victor Botev said.

It's not going to be just, let's say, two times cheaper, but multiple times cheaper, if it shows the results.

However, based on Google's benchmark result on its new Gemma models, Iris plans to evaluate how the 7B model compares with the Llama 13 billion model. Gemma 7B will enable Iris to use only one GPU if it performs as promised, driving the cost down for the startup.

"Anything above that requires you to actually set up multiple GPUs, which is not easy to operate," Botev added. Working with multiple GPUs requires special infrastructures, libraries and software stacks.

"It's not going to be just, let's say, two times cheaper, but multiple times cheaper, if it shows the results," Botev said. "It can fit into one GPU, it can allow us to generate multiple answers because the memory will be enough, and that will drive the cost significantly ... if the reasoning capabilities are exactly as they say, and if it's easy enough to fine-tune and modify for [our] particular use case."

The Gemma 2B model, by comparison, seems similar to the Llama 7 billion model, which does not have good reasoning capabilities, according to Botev.

"That would be interesting the moment they actually reach the reasoning capabilities of the current 13 billion-parameter models," he said.

Open model and responsible AI

While Gemma is an open model that limits what companies such as Iris can do with it, it still allows for experimentation, Botev added.

Google has specifically positioned Gemma as an "open" rather than "open source" model. It essentially is a type of open source model that does not include an open source license. As such, it restricts certain ways users can deploy it, especially those Google deems unsafe or not in accordance with responsible AI standards.

Open models allow free access to various model sizes, redistribution and variant ownership, and users can create and publish different model variants, the cloud vendor said.

However, Google is being deliberate in how it says it is rolling out the technology responsibly.

"They're using some of the best-in-class responsible design techniques to build this thing," Beccue said.

For example, Gemma incorporates automated techniques that filter out personally identifiable information. It also uses reinforcement learning from human feedback to combat bias. Google also employed manual red teaming efforts and automated advanced adversarial testing for the models.

"[The models] feel very well thought through in terms of being able to rely on them," Beccue said. "This is a pretty cool move, and I think this is going to be very interesting."

Esther Ajao is a TechTarget Editorial news writer covering artificial intelligence software and systems.