Google AI—a division of the search giant—unveiled Gemma today, a new family of open-source language models derived from its recently released Gemini suite of AI tools. With Gemma, Google is directly challenging the leading position of LLMs based on Meta’s LLaMa and Mistral.
“We have a long history of supporting responsible open source and science, which can drive rapid research progress,” wrote Demis Hassabis, co-founder of Google DeepMind, in a tweet. “So we’re proud to release Gemma: a set of lightweight open models, best-in-class for their size, inspired by the same tech used for Gemini.”
Gemini is the the closed-source, commercial, multimodal LLMs recently released by Google to replace PaLM-2 in powering its suite of online services, competing directly against OpenAI’s ChatGPT Plus.
Gemma is available in two versions: Gemma 2B, a lightweight model trained on two billion parameters that is optimized for efficiency and can be run in a wider spectrum of hardware, and Gemma 7B, a more robust model trained on 7 billion parameters, which provides better results but requires more computing power.
“At Google, we believe in making AI helpful for everyone,” Google said in an official announcement, emphasizing its vision of democratizing AI technology.
Releasing an open source model may more immediately impact AI developers and software engineers, but the benefits reach end users after they have been tested, fine tuned, and built into products and services.
“As we continue to expand the Gemma model family, we look forward to introducing new variants for diverse applications,” the company added.
Google’s announcement covered the technical and philosophical foundations of Gemma, saying the model “is designed with our AI Principles at the forefront.”
“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,” it said.
Shortly after its release, Gemma quickly climbed the ranks on the oft-referenced Hugging Face leaderboard, reaching to the second spot in the trending list and beating all other open source LLMs and AI models—except for Stability AI’s Open Cascade, a new text-to-image generative AI.
Real numbers back Gemma’s claimed prowess: it has demonstrated superior performance over Meta’s LlaMa-2 across several synthetic benchmarks. For example, Gemma outperforms LlaMa-2 in tests like MMLU (knowledge covering 57 scientific disciplines), Hellaswag (evaluating basic reasoning in everyday tasks), MATH (assessing mathematical capabilities), and HumanEval (testing coding proficiency).
Developers interested in testing Gemma can do so for free using Kaggle, a data science and artificial intelligence platform that Google built. The company has also made Gemma available via Google Colab notebooks, and is giving $300 in credits for first-time Google Cloud users.
“Researchers can also apply for Google Cloud credits of up to $500,000 to accelerate their projects,” Google said.
Edited by Ryan Ozawa.
