Microsoft released its PHI-4 Artificial Intelligence (AI) model on Friday. The company’s latest small language model (SLM) joins its open source PHI basic model family. The AI model was eight months after the release of PHI-3 and eight months after the introduction of the PHI-3.5 series AI model. The tech giant claims that SLM is more capable of solving complex inference-based queries in fields such as mathematics. Furthermore, it is said to perform well in conventional language processing.
Microsoft’s PHI-4 AI model is available by embracing faces
Each PHI series has been launched in a mini variant so far, however, no mini model is accompanied by the PHI-4 model. Microsoft emphasized in a blog post that PHI-4 is currently available on Azure AI Foundry under the Microsoft Research License Agreement (MSRLA). The company plans to be available on next week’s hug faces.
The company also shared benchmark scores through its internal tests. Based on these, the new AI model can significantly upgrade the functions of the older generation model. The tech giant claims that the PHI-4 is better than the larger model of the Mathematics Competition Pro 1.5. It also publishes detailed benchmark performance in a technical paper published in the online journal Arxiv.
On the security front, Microsoft says Azure AI Foundry has a range of features that help organizations measure, mitigate and manage AI risks throughout the development lifecycle for traditional machine learning and generated AI applications. Additionally, enterprise users can use Azure AI content security features such as timely shields, ground detection and other content as content filters.
Developers can also add these security features to their applications through a single application programming interface (API). The platform monitors application quality and security, adversarial and timely attacks, and data integrity, and provides developers with real-time alerts. PHI users who access it through Azure will use this feature.
It is worth noting that smaller language models are often trained after deploying synthetic data, allowing them to quickly gain more knowledge and be more efficient. However, in the real world, post-training results are not always consistent.
