Microsoft introduces the Phi-2 model with 2.7 billion parameters that outperforms many others large language models (LLMs), which are up to 25 times larger.

There are always new developments and improvements in the world of AI. For example, Microsoft is now introducing the Phi 2 model, which is said to offer better performance than other LLMs. The model is trained based on this Textbook qualitywhere reasoning, knowledge and common sense take center stage.

Phi-2 outperforms other LLMs

The performance of LLMs is usually measured by their overall size, expressed in parameters. The more parameters, the more capabilities. Phi-2 breaks this rule and impresses with enhanced capabilities without following this rule.

One of the most striking statements is undoubtedly that the model can even exceed the performance of Google LLC’s Gemini Nano, which was introduced last week. Gemini Nano is the most efficient model in the Gemini series and can also run on smartphones.

successes due to Textbook quality

The Phi-2 model is trained based on a Textbook quality-Record. Such a sentence is carefully constructed from data selected to convey reasoning, knowledge and common sense. This allows the model to learn more from less information.

Phi-2 already knew several predecessors. Earlier this year, Phi-1 was launched with 1.3 billion parameters fine-tuned for basic Python coding tasks. Phi-1.5 followed in September with 1.3 billion parameters, trained on new data sources containing various synthetic texts generated using natural language programming.

The model has been tested multiple times by Microsoft researchers for challenges in language comprehension, reasoning, mathematics and coding. According to Microsoft, Phi-2 is the ideal platform for researchers who want to delve deeper into improving AI safety, interpretability and ethical development.