Microsoft has finally made its new tiny tongue model Phi-4 available for free on the Hugging Face platform. This model with 14 billion parameters can now be downloaded, optimized and used for free.
Why is this important?
Despite its small size, Phi-4 outperformed OpenAI’s Llama 3.3 model with 70 billion parameters (which is roughly five times larger) and OpenAI’s GPT-4o Mini model in several tests. In fact, the Phi-4 model outperformed the Gemini 1.5 Pro and the smaller version of the GPT-4o in math tests.
In a technical article, Microsoft explained the various techniques and high-quality data used to train this model. This model is known for its strong capabilities in solving complex problems.
Harkirat Behl, one of the creators of this model, stated in an interview: “Large models are trained on very diverse data and store information that may not be relevant.” With enough effort to select high-quality data, the performance of large models can be achieved or even surpassed, he added.
In the development of Phi-4, Microsoft did not particularly focus on optimizations related to inference and focused more on artificial data. Harkirat Behl revealed that by publishing the architecture of the model, developers can further optimize it and run it for local use on PCs and laptops by reducing the size of the model.
After Meta, Microsoft is one of the big companies that has made significant progress in the field of open source language models. The previous version of this model, Phi-3.5, was also released for free on Hugging Face.
However, in the competition of open source models, even Microsoft and Meta are not leading. Currently, the DeepSeek-V3 model from China is on top. This very large model with 671 billion parameters outperforms the Llama 3.1 meta model with 405 billion parameters and many other closed models. Also, this Chinese model is three times faster than its previous version, DeepSeek V2.
Phi-4 supports ten Indian languages, Behl said. “I personally put a lot of effort into making Phi-4 able to interpret ten common Indian languages,” he added. This shows that Microsoft has invested heavily in India.
RCO NEWS