Microsoft He introduced a new class of highly efficient AI models that process the text, images and speech at the same time, requiring much less computing power than existing systems. New Phi-4 modelsPublished today, they represent a breakthrough in the development of small language models (SLM), which give possibilities previously reserved for much larger AI systems.
Phi-4-Multimodalmodel with only 5.6 billion parameters and PH-4-MiniWith 3.8 billion parameters, they exceed competitors of comparable size, and even match or exceed the performance of models twice their size in some tasks, in response to Microsoft Technical report.
“These models are aimed at strengthening advanced programmers of AI’s capabilities,” said Weizhu Chen, AI generative vice chairman at Microsoft. “Phi-4-Multimodal, with its ability to process speech, vision and text at the same time, opens new possibilities of creating innovative and contextual applications.”
Technical achievement appears at a time when enterprises are increasingly looking for AI models that may operate on standard equipment or on “edge” – directly on devices than in the cloud data centers – to scale back costs and delays while maintaining data privacy.
How Microsoft has built a small AI model that does all this
What he sets Phi-4-Multimodal Besides, his novel “LORAS mixture“The technique, enabling it to support text, images and speech in one model.
“Using the LORAS mixture, Phi-4-Multimodal expands multimodal capabilities, while minimizing the disturbances between methods”, Research article countries. “This approach allows trouble -free integration and ensures consistent performance in tasks related to text, images and speech/sound.”
Innovation allows the model to take care of its strong language capabilities while adding a vision and speech recognition without performance degradation, which frequently occurs when models are adapted to many input types.
The model has won the highest position in Hugging the face of the OpenSr Plaque of Leaders with a word error indicator of 6.14%, exceeding specialized speech recognition systems comparable to Whisperv3. It also shows competitive results in vision tasks, comparable to mathematical and scientific reasoning with images.
Compact artificial intelligence, huge influence: Phi-4-Mini sets new performance standards
Despite the compact size, PH-4-Mini It shows unique possibilities in text tasks. Microsoft reports that the model “exceeds models of similar sizes and is in the scope of two -larger models” in various comparative tests.
The performance of the model in mathematical and encoding tasks is particularly noteworthy. According to Research article“Phi-4-mini consists of 32 layers of transformers with a hidden state 3 072” and comprises the attention of the group’s inquiry to optimize memory consumption to generate long contact.
On Benchmark GSM-8K MATHPhi-4-Mini achieved the results of 88.6%, exceeding most models of 8 million parameters, while at the mathematical level he reached 64%, much higher than competitors of comparable size.
“In the case of a mathematical reference point, the model exceeds models of similar sizes with large margins, sometimes over 20 points. It even exceeds the models’ results twice as much, “notes the technical report.
Transformational implementation: actual Phi-4 performance in operation
CapacityAI response engine, which helps organizations unite a variety of information sets, has already used the Phi family to extend the performance and accuracy of the platform.
Steve Frederickson, product head with capability, said in statement“From our initial experiments, what really impressed us on Phi was his extraordinary accuracy and ease of implementation, even before adaptation. Since then, we were able to increase both accuracy and reliability, while while maintaining profitability and scalability, which we valued from the very beginning. “
The capability has reported 4.2 -costs of costs in comparison with competitive work flows while reaching the same or higher quality results for preliminary tasks.
AI without restrictions: Microsoft Phi-4 models bring advanced intelligence anywhere
For years, the development of artificial intelligence has been powered by peculiar philosophy: Bigger is higher. More parameters, larger models, greater calculation requirements. But Microsoft Phi-4 models query this assumption, proving that power is not only about the scale-it involves performance.
Phi-4-Multimodal AND PH-4-Mini They are designed not for technological giant data centers, but for the real world – where the computing power is limited, the fears of privacy are the most vital, and AI must act without any problems without continuous connection with the cloud. These models are small, but they have weight. Phi-4-Multimodal integrates speech, vision and processing of text with a single system without dedicating accuracy, while Phi-4-Mini provides mathematics, coding and reasoning on couples with models twice as high.
It’s not only about increasing AI efficiency; It’s about making him more accessible. Microsoft has placed PHI-4 for universal adoption, because of which it is available Azure ai foundryIN Huggingand NVIDIA API directory. The goal is clear: AI, which is not blocked by expensive equipment or huge infrastructure, but one that may work on standard devices, on the fringe of the network, and in industries where computing power is rare.
Masaya Nishimaki, director of the Japanese company AI Headwaters Co., Ltd., sees first -hand influence. “Edge AI has exceptional performance even in environments with unstable network connections or if confidentiality is the most important,” said Wa statement. This means artificial intelligence, which may function in factories, hospitals, autonomous vehicles-members, in which real-time intelligence is required, but in the case when traditional cloud-based models are not short.
At the heart of Phi-4 is a change in considering. AI is not only a tool for people with the largest servers and the deepest pockets. It is an ability that, if well designed, can work anywhere, for everyone. The most revolutionary thing in Phi-4 is not what it will probably do-this is where it will probably do it.