Microsoft’s new Phi-4 AI models pack big performance in small packages

Microsoft Text, images and speeches are simultaneously presented high efficiency AI models at the same time, relative to the existing systems. New Phi-4 modelsToday shows a progress in the development of small language models (SLMS), which is released, greater protected opportunities for larger AI systems.

Phi-4-Multimodala model with a total of 5.6 billion parameters and PH-4-miniWith 3.8 billion parameters, make similar sizes, and even according to Microsoft, determine or exceed the performance of models of models twice in the deductible tasks Technical report.

“These models develop developers are designed to develop developers,” said Weizhu Chen, Senior Vice President, Meshirchi Ai. “PHI-4-Multimodal opens new opportunities to create innovative and contextual applications with the ability to process speech, vision and text.”

Technical achievements are generated in an increasingly standard AI models that can work on a standard apparatus or “edge“- Cloud data centers, not direct devices – reduce costs and delays while maintaining the confidentiality of data.

How Microsoft has built a small AI model that makes it all

What does it build Phi-4-Multimodal is a separate novel “a mixture of lauras“The technician allows you to manage text, pictures and speech entries within a model.

“Using a mixture of Loras, PHI-4-Multimodal, expands multimodal capabilities while minimizing the intervention between minibuses” Research paper States. “This approach provides seamless integration and provides consistent performance between text, images and speech / audio tasks.”

When the innovation is often adapted for the entries of the model, it allows you to maintain strong language opportunities during vision and speech recognition without frequent performance degradation.

The model claimed to be in the upper position Face OpenAsr Leaderboard 6.14% with a word error rate, best speech recognition systems Mental. It also demonstrates a competitive performance on visual assignments such as mathematical and scientific justification.

Compact AI, mass effect: PHI-4-MINI sets new performance standards

Despite the compact size, PH-4-mini Demonstrates exceptional opportunities in text-based positions. Microsoft notifies the model, “Models of similar sizes and are larger than various languages.”

It is especially noteworthy, performance on the model math and coding tasks. According to Research paper“The PHI-4-mini consists of 32 transformer strata with a secret state volume and focuses on a group request to optimize memory usage for a long-term generation.

On GSM-8K Math BenchmarkPHI-4-mini, 88.6%, according to the math criterion, won 88.6% of the most than 8 billion paramels, which reached 64% higher than similar sizes.

“Math Benchmark, sometimes the scores of similar sizes, which are more than 20 point margins, sometimes twice larger models scores,” the technical report records.

Transformative Accommodation: Real world efficiency in the activities of PHI-4

PowerThe AI ​​response engine that helps organizations connect various data networks has already used the PHI family to increase the efficiency and accuracy of the platform.

Steve Frederickson, the head of the product, told a statement“Of course, what is affected by our preliminary practices, the thing affected by the PHI, we have been able to increase both accuracy and reliability, both accuracy and reliability and reliability, as well as the effectiveness and size of the beginning.”

Potension saving 4.2x costs compete compared to the same or higher quality results for tasks to pre-processing the tasks.

AI Unlimited: Microsoft’s Phi-4 models bring advanced intelligence everywhere

AI development has been managed by a single philosophy for years: better. The more parameters, larger models, larger calculation requirements. However, Microsoft’s Phi-4 models prove that this hypothesis is about its effectiveness, which proves that the power is not only scale.

Phi-4-Multimodal and PH-4-mini Technology giants are not for data centers, but for the real world – privacy concerns in places where the calculation power must work without constant contact with paramount and cloud. These models are small, but they carry weight. PHI-4-Multimodal, Accuracy, PHI-4-mini, Phi-4-mini, drawing, coding and substantive performance combines a system, vision and text processing twice.

This is not just to make the EU more efficient; This is to make it more accessible. Microsoft has placed the Phi-4 to take it wide Casting AZURE AI, Hug faceand and and Nvidia API Catalog. The goal is clear: where it is not locked behind expensive apparatus or mass infrastructure, but in standard devices, where something that can operate in the edges of the networks and in the settlements.

Japan’s company AI Firm Headwaters Co., Ltda, a director, dessert in a director, sees his effect in advance. “EDGE AI also shows outstanding performance in environments with unstable network connections or the principal is the basis of the privacy,” he said statement. This means that in factories, hospitals, which can operate in autonomous vehicles, where real-time intelligence is required, but where traditional cloud-based models are short.

Represents PHI-4 Thinking Change in Core. AI is not a tool for those with the biggest servers and the deepest pockets. If it is well prepared, it is an ability to work anywhere, for anyone. The most revolutionary thing about Phi-4 is not – it’s where it can do it.