Gemma: Google's Open-ish LLMs

Google may have been a bit late to the AI party when it comes to releasing Large Language Models (LLMs), with OpenAI’s ChatGPT and Meta AI’s open source LLaMA beating them to the punch. But after playing catch-up with their release of Bard, and more recently Gemini, they may have finally come close to matching their competitors in terms of capabilities.

This is not to say that everything is coming up roses for Google now — they have come under intense criticism for what many see as excessive guardrails that have been built into Gemini. The tool will often refuse to answer seemingly innocent requests, citing “safety” concerns. More recently, Google took Gemini’s image generation capabilities offline in response to revelations that the tool produces inaccurate results, seemingly as a result of Google’s intentional manipulation of the model.

It is within this backdrop that Google has taken their next step into the booming generative AI market. They have just announced the release of a new family of models called Gemma. These models are lightweight versions of their flagship Gemini model, and were built from the same research and technology. At present, there are two Gemma models available — one with two billion parameters, and another with seven billion.

The Gemma models are not exactly open source, but instead are what Google calls “open models.” In a nutshell, that means that the model weights are freely available for anyone to access, but the terms of use, redistribution, and so on, vary by model and may not look a lot like what one would expect of an open source project. Also of note is Google’s clause that users must “agree to avoid harmful uses.” While this seems like a prudent measure to take on the surface, it is also very subjective — one need look no further than Gemini’s own sometimes frustrating refusal to answer harmless prompts to understand that. As this clause would seem to give Google the ability to shut down any use of the model that they do not like, it will undoubtedly deter many users from utilizing the technology.

Despite these concerns, Gemma exhibits excellent performance and efficiency, which could still draw many users to it. The models are capable of running locally on a typical desktop or laptop computer, and benchmarks demonstrate that they achieve state of the art performance when compared with other existing options available today.

A quick start guide and instructions to access the model are now available for those that want to give Gemma a whirl. It should be simple to get started — toolchains have been released for popular frameworks like PyTorch and TensorFlow, and ready-to-use Google Colab notebooks can help even beginners to learn the basics.

Looking ahead, Google plans to introduce more models in the Gemma family under the same “open model” paradigm. These new models will be tuned to specific applications, so if you do not see what you need yet, be sure to check back later for updates.