Microsoft issued a document: Phi-2 AI model performance surpasses Google Gemini Nano-2, with parameter size reaching 2.7 billion

News on December 13, Microsoft issued a press release today, stating that its Phi-2 2.7B model is superior to the Gemini Nano-2 3.2B released by Google in many aspects.

Phi-2 2.7B model

According to reports from this site in November last year, at the Ignite 2023 conference, Microsoft released Phi-2 with parameters reaching 2.7 billion and a performance ratio of The previous version has been significantly improved

Microsoft released Phi-1 in June this year, with only 1.3 billion parameters, suitable for QA Q&A, chat format and code and other scenarios. The model is trained entirely on high-quality data and outperforms competing models on benchmarks by up to 10x.

Microsoft updated the Phi-1.5 version in September this year, which also has 1.3 billion parameters. This version can be used for writing poetry, writing emails, writing stories, and summarizing texts. In benchmarks of common sense, language understanding, and reasoning, the model was able to keep up with models with up to 10 billion parameters in some areas

Microsoft's latest release of Phi-2 now has 2.7 billion parameters, despite its scale Double that of previous versions, but still smaller than other mainstream language models

Microsoft says Phi-2 shows significant improvements in logical reasoning and security Improve. With the right fine-tuning and customization, small language models are powerful tools for cloud and edge applications.

Google Gemini Nano-2 3.2B

Gemini Nano is a model version specially built to run natively on small devices. The latest 2.0 version has 3.2 billion parameters and will be the first to be equipped on Pixel 8 On the Pro model.

In the recording scene, select the recording file and click the "Transcript" tab, then click the "Summary" button at the top, the App will generate multiple transcripts Points related to the content of the recording.

In the Gboard input method, Gemini Nano will achieve "high-quality conversation-aware replies". WhatsApp will be the first App to support smart replies, and other Apps will also gain support next year.

Comparison

In its latest blog post, Microsoft compared the Phi-2 model with Google’s Gemini Nano-2 model, saying that many of the performances of Phi-2 are better than the Gemini Nano-2 model.

Microsoft also pointed out that the performance of Phi-2 has exceeded Llama-2's 7 billion parameters and 13 billion parameters, as well as Mistral's 7 billion parameters

############################################################# #, interested users can read in depth. ###

The above is the detailed content of Microsoft issued a document: Phi-2 AI model performance surpasses Google Gemini Nano-2, with parameter size reaching 2.7 billion. For more information, please follow other related articles on the PHP Chinese website!