Become a member of our daily and weekly newsletters for the latest updates and exclusive content about leading AI coverage. Leather
Microsoft Has introduced a new class of highly efficient AI models that process text, images and speech at the same time and at the same time require considerably less computing power than other available systems. The new Phi–4 modelsRepleated today, represent a breakthrough in the development of small language models (SLMs) those possibilities that have previously been reserved for much larger AI systems.
Phi–4–multimodalA model with only 5.6 billion parameters, and Phi-4-MiniWith 3.8 billion parameters, perform better than competitors of similar size and at certain tasks even correspond to the performance of models twice their size, according to Microsoft technical report.
“These models are designed to enable developers advanced AI options,” said Weizhu Chen, vice -president, generative AI at Microsoft. “Phi-4-multimodal, with its ability to process speech, vision and text at the same time, opens new possibilities for creating innovative and context conscious applications.”
This technical performance comes at a time when companies are increasingly looking for AI models that can be performed on standard hardware or on the “lead” – directly on devices instead of cloud data centers – to reduce costs and latency while retaining data privacy.
How Microsoft built a small AI model that it all does
What stucks Phi-4-multimodal Apart from the novel ‘Mix from Loras“Technology, so that it can process text, images and speech entry into a single model.
“By using the Loras mixture, PHI-4-Multimodal expands multimodal possibilities, while the interference between modalities is minimized” research paper States. “This approach makes seamless integration possible and ensures consistent performance between tasks with text, images and speech/audio.”
The innovation enables the model to maintain its strong language possibilities and at the same time add vision and speech recognition without the performancegradation that often occurs when models are adapted for multiple input types.
The model has the top position on the Hugging face open axle leaderboard With a word error percentage of 6.14%, it performs better than specialized speech recognition systems such as Whisperv3. It also shows competitive performance on vision tasks such as mathematical and scientific reasoning with images.
Compact AI, Massive Impact: Phi-4-Mini sets new performance standards
Despite its compact size, Phi-4-Mini Shows exceptional possibilities on text -based tasks. Microsoft reports that the model ‘performs better than models of similar size and is on the par with models twice [as large]’About different languages-making benchmarks.
The performance of the model about mathematics and coding tasks are particularly remarkable. According to the research paper“Phi-4-Mini consists of 32 transformer layers with a hidden state size of 3,072” and contains the group’s query’s attention to optimize memory use to generate a long context.
On the GSM-8K Mathematical BenchmarkPHI-4-Mini achieved a score of 88.6%, which performed better than most models of 8 billion parameters, while reaching 64% at the mathematical benchmark, considerably higher than competitors of comparable size.
“For the mathematical benchmark, the model performs better than models of similar size with large margins, sometimes more than 20 points. It performs even better than twice as long as models scores, “notes the technical report.
Transforming implementations: PHI-4’s Real-World Efficiency in action
CapacityAn AI “answer motor” that helps organizations to unite different data sets has already used the Phi family to improve the efficiency and accuracy of the platform.
Steve Frederickson, head of the product on capacity, said in one rack“Of our first experiments, what really made an impression on us about the Phi the remarkable accuracy and the convenience of commitment, even before adjustment. Since then we have been able to improve both accuracy and reliability, while maintaining the cost -effectiveness and scalability that we have rated from the start. “
Capacity reported a 4.2x cost savings compared to competing workflows and at the same time achieving the same or better qualitative results for pre -process tasks.
AI without limits: bring Microsoft’s Phi-4 models everywhere bring advanced intelligence
For years, AI development is powered by a single philosophy: larger is better – more parameters, larger models, larger computational requirements. But the Phi-4 models from Microsoft days that assumption, which show that strength is not only about scale is about efficiency.
Phi-4-multimodal And Phi-4-Mini Are not designed for the data centers of technical giants, but for the real world – where the computing power is limited, privacy problems are of the utmost importance and AI must work seamlessly without a constant connection with the cloud. These models are small, but they have weight. PHI-4-multimodal integrates speech, vision and word processing into a single system without sacrificing accuracy, while Phi-4-Mini delivers mathematics, coding and reasoning performance with models twice the size.
This is not just about making AI more efficient; The point is to make it more accessible. Microsoft has positioned PHI-4 for widespread adoption, making it available via Azure ai Foundry” Hug and the NVIDIA API -Catalogus. The goal is clear: AI that is not locked behind expensive hardware or massive infrastructure, but can previously work on standard devices, on the edge of networks and in industries where the computing power is scarce.
Masaya Nishimaki, a director of the Japanese AI company Headwaters Co., Ltd., sees the impact firsthand. “Edge AI shows excellent performance, even in environments with unstable network connections or where confidentiality is of the utmost importance,” he said in a rack. That means AI that can function in factories, hospitals, autonomous vehicles places where real-time intelligence is required, but where traditional cloud-based models fail.
In the core, Phi-4 represents a shift in thinking. AI is not only a tool for people with the largest servers and the deepest bags. It is a possibility that, if well -designed, can work for everyone everywhere. The most revolutionary thing about Phi-4 is not what it can do is where it can do it.
Source link
