Continuously update itsPhi Natural Language ModelLater, Microsoft announced the launch of a new parameter scale of 140 billion groupsLightweight natural language model Phi-4, boasting performance comparable to large natural language models including Llama.
Phi-4 emphasizes the ability to improve mathematical reasoning and language comprehension within a limited parameter scale, and can even surpass the mathematical processing performance of the 700 billion parameter scale version of Llama.
Phi-4 primarily utilizes high-quality synthetic data combined with screened real-world data, and employs recalibrated post-training technology to significantly improve inference accuracy. It also improves operational efficiency through a smaller parameter scale, making it easily deployable on more terminal devices.
Microsoft has already made Phi-4 available through the Azure AI Foundry platform, and will subsequently offer it through the Hugging Face platform for hosting, thereby attracting more companies and developers to use it and allowing Phi-4 to be widely used in more computing scenarios.
