The open source big model track welcomes a heavyweight new product.
On February 21st local time, Google announced the official launch of a new open source big language model (LLM) called Gemma, aimed at helping developers and researchers responsibly build artificial intelligence.
It is reported that the Gemma big model shares technology and infrastructure with Google's largest and most powerful artificial intelligence model, Gemini. "Inspired by Gemini, Google DeepMind collaborated with other Google teams to develop Gemma, which is named after Gemma, meaning 'gem' in Latin."
However, compared to Gemini, Gemma is more lightweight. Meanwhile, Gemma remains free to use, its model weights are also open-source, and commercial use is allowed.
Google has released two models with different weight scales, Gemma 2B (2 billion parameters) and Gemma 7B (7 billion parameters). Each scale has pre trained and instruction fine-tuning versions, allowing all organizations (regardless of size) to responsibly conduct commercial and distribution.
On the same day that Google released Gemma, the popular chip manufacturer Nvidia also announced a partnership with Google to ensure the smooth operation of the Gemma model on its chips. Nvidia also stated that its chatbot software Chat With RTX will soon support Gemma.
It is worth noting that Google also emphasizes that Gemma can surpass larger models on key benchmarks. What's even more impressive is that Google Gemma can run on laptops.
Google has stated that Gemini is the largest and most powerful AI model widely used today. Compared to other open models, Gemma 2B and 7B can achieve the best performance in their class within their scope. The Gemma model can run directly on developers' laptops or desktops, "It's worth noting that Gemma surpasses larger models on key benchmarks while adhering to our strict standards of safe and responsible output."
Along with the open source model, Google also released a technical report on Gemma's performance, dataset composition, and modeling methods in detail. Researchers have found in a technical report that Gemma supports a vocabulary size of 256K, which means it can provide better and faster support for languages other than English.
Comparison of Llama 2 parameters released by Gemma and Meta, from Google's official website
Gemma was also launched as soon as possible on the well-known open-source model libraries HuggingFace and HuggingChat. Shortly after its launch, both Gemma 2B and 7B models have reached the top of HuggingFace's "Big Language Model List".
AI industry expert and author of the deep learning framework Keras, Franois Chollet, further stated that the position of the strongest open source big model has now changed ownership.
Gemma's competitor Llama 3 is also about to be released. On January 19th, Meta co-founder and CEO Zuckerberg announced that Meta is training Llama 3 and will continue to open source in a responsible manner.