Here is the package
Microsoft has unveiled Phi-4, the latest in its Phi series of generative artificial intelligence (AI) models. The streamlined architecture includes advances in matic problem solving.
The new model, which has 14 billion parameters, reportedly aims to compete with other built-in AI models such as GPT-4o Mini, Gemini 2.0 Flash, and Claude 3.5 Haiku.
According to a blog post, Phi-4 is available with limited access through Microsoft's Azure AI Foundry platform and is restricted to research purposes under the Microsoft Research License Agreement.
Phi-4: Improving performance in matic inference
Microsoft has positioned Phi-4 as a leader in matic problem solving, citing significant performance gains over its predecessors and similar models. The company touted the capabilities of its AI model after Phi-4 achieved top scores on several benchmarks.
On the GPQA test, it scored 56.1, beating GPT-4o’s 40.9 and Llama-3’s 49.1. On the MATH benchmark, Phi-4 scored 80.4, reflecting its advanced capabilities in tackling complex matic problems. It also excelled on programming benchmarks, earning a score of 82.6 on HumanEval.
Additionally, Phi-4 has demonstrated prowess in real-world scenarios, including high scores on problems from the American Mathematical Association's AMC-10/12 competitions. These results point to potential applications in scientific research, engineering, and financial modeling, areas where accuracy and mathematical reasoning are critical.
While larger models like OpenAI's GPT-4o and Google's Gemini Ultra operate with hundreds of billions or even trillions of parameters, Phi-4 shows that smaller, simplified architectures can achieve superior performance on specialized tasks.
Microsoft attributes the progress of Phi-4 to the integration of high-quality synthetic data alongside human-generated content datasets, as well as unannounced improvements made during the post-training phase. These efforts reflect a broader trend in the AI industry, where research teams are increasingly focusing on innovations in the use of synthetic data and post-training improvements.
Alexander Wang, CEO of Scale AI, recently highlighted this shift, noting that the industry has reached a “pre-training data wall,” adding that companies will now race to develop more efficient AI models.
Microsoft continues to emphasize responsible development of AI solutions, incorporating robust safety measures into Phi-4 and its predecessors. With Azure AI Foundry, users have access to tools designed to assess and mitigate risks across the AI development lifecycle.
These tools include instant shields, which protect against inappropriate or harmful inputs, protected material detection (sensitive content in outputs), and grounding detection to ensure that outputs are accurate and relevant to reality.
Additionally, there are built-in features in the Azure AI Content Safety Toolkit, allowing developers to apply filters and monitor applications to ensure quality, safety, and data integrity. Real-time alerts provide timely interventions to address issues such as hostile claims and content deviations.
Azure AI Foundry also evaluates iterative models using built-in and custom metrics, giving developers the flexibility to tune AI applications for optimal performance.