Microsoft opened its PHI-4 small language model on Wednesday. The latest artificial intelligence (AI) model in the PHI family was released last month, but was only available through the company’s Azure AI foundry at the time. At the time, the Redmond-based tech giant said it would soon produce source code for AI models available in the public domain. Now, those interested can access reasoning-focused AI models by embracing faces. Microsoft also uses the model for academic and commercial use cases.
Microsoft Open Source PHI-4 AI Model
Shital Shah (Shital Shah) of Microsoft AI technician announced the weight of the PHI-4 AI model on the embracing face announcement X (formerly known as Twitter). The AI model can comply with MIT academic and commercial use licenses. Those interested can access the model list here.
It is said that SLM was launched eight months after the release of the PHI-3 AI model, and could make significant improvements in major reasoning-based queries in fields such as mathematics. The context window for PHI-4 is 16,000 tokens and is trained on a 9.8 trillion token dataset.
Embrace Faces lists the source of training data, lists the dataset that includes publicly available high-quality education data and code, synthetic data across a wide range of topics, obtained academic books and question answer datasets, and chat formats Supervise data.
It is worth noting that this is a text-only model, meaning it only accepts text as input and output. The AI model has 14 billion parameters. Microsoft notes that the AI model is built on an intensive decoder transformer architecture.
At the time of release, Microsoft also shared benchmark scores for the AI model. Based on them, the company claims that the latest iteration of PHI SLM is better than the Gemini 1.5 Pro model in the math competition problem benchmark.
The PHI-4 AI model is also accessible via Microsoft’s Azure AI Foundry. The platform also provides helps developers and businesses manage AI risks. It also has functions such as timely shield, ground detection and content filters. These security features can also be exported to applications using the company’s application programming interface (API).
