The open source field of big models has welcomed heavyweight new players.
On Wednesday (February 21) local time, technology giant Google released a new generation of open source model "Gemma", which the company claims is the "most advanced" open model series among lightweight models, surpassing the originally strongest Mistral 7B.
According to Google's official website, Gemma means "gem" in Latin, and the open model series Gemma named after it was developed by a team such as Google DeepMind, using the same research and technology as creating Gemini models.
At present, there are two weight scale models in the Gemma series, namely "Gemma 2B" and "Gemma 7B", with 2 billion parameters and 7 billion parameters, to meet the different needs of developers.
In terms of performance, the official page shows that Gemma 7B has a score rate of 64.3% in MMLU (Massive Multi Task Language Understanding). In the same scale, the originally strongest Mistral 7B has a scoring rate of 62.5%, and Meta's Llama-2 7B and 13B also have a certain gap with it.
Gemma also launched HuggingFace and HuggingChat as soon as possible, and everyone can try out its generation ability:
On the centralized leaderboard of Hugging Face's LLM, Gemma's 2B and 7B models also have average scores at the top of their respective parameter scales.
Google wrote in a press release that Gemma 2B and 7B exhibit state-of-the-art performance compared to other open models at the same parameter scale, even surpassing some larger models, thanks to sharing technology and infrastructure with the company's strongest AI model, Gemini.
Software engineer Jeff Dean stated that despite its small size, the Gemma model has clearly surpassed larger models in key benchmark tests, including the Llama-2 7B and 13B, as well as the trendy Mistral 7B.
And Gemma can run directly on developers' laptops or desktops. In addition to lightweight models, Google has also launched tools to encourage collaboration and guidelines for responsible use of these models.
Keras author Francois Chollet directly stated that the location of the strongest open source big model has now changed ownership.
Google and Alphabet CEO Sundar Pichai stated that Gemma has demonstrated powerful performance, and starting today, the model will be available globally and can run on laptops, workstations, or Google Cloud.
Demis Hassabis, CEO of Google DeepMind, stated that the lightweight and open-source Gemma is the best performing model in its class size.
Nowadays, Google is not only targeting OpenAI as its competitor, but also intends to dominate everything, with the new model "far ahead" of Zuckerberg's Llama-2.
Google plans to launch more Gemma variants in the future to adapt to a wider range of application scenarios. Currently, developers who want to start using these models can access them through the data science platform Kaggle, Google's Colab laptop, or through Google Cloud.
Daily Economic News Comprehensive Google Official Website and Public Information