Microsoft Unveils Phi-3.5 AI Models, Challenging Industry Giants

Microsoft releases new series of Phi-3.5 models that beat out rivals

Microsoft has released the latest batch of open-source small AI models in the Phi series called Phi-3.5. The company has claimed that the three small language models beat competitors including Google's Gemini 1.5 Flash, Meta's Llama 3.1, and even OpenAI's GPT-4o in some benchmarks. The set of three new Phi-3.5 models include the 3.82 billion parameter Phi-3.5-mini-instruct, the 41.9 billion parameter Phi-3.5-MoE-instruct, and the 4.15 billion parameter Phi-3.5-vision-instruct, which were all designed for tasks like basic and fast reasoning, more powerful reasoning and vision tasks like image and video analysis, respectively. All of these three models are available for download for free and can be run using a local tool like Ollama. Despite its tiny size, the Phi-3.5 Mini Instruct model can process images as well as text and is multilingual too. (For top technology news of the day, subscribe to our tech newsletter Today's Cache) The model performed quite well at the reasoning tasks and was only beaten by GPT-4o-mini out of its rivals. It also performed well on math benchmarks, significantly passing other models like Llama and Gemini. The Phi-3.5-MoE model is the first model in this class and can combine multiple different model types into one, so each specialises in different tasks. The model was able to beat Gemini Flash 1.5, which is used in the free version of the Gemini chatbot. Microsoft revamps reporting on business units, offers clarity on AI benefits It also has a comfortably large 128k context window which although significantly smaller than Gemini is equal to ChatGPT and Claude. While the smallest model in the range was trained on 3.4 trillion tokens of data using 512 Nvidia H100 GPUs over 10 days, the MoE model used 4.9 trillion tokens and was trained over 23 days. The major advantage of using small language models in this size is that they can be used in an app or even installed on IoT device easily because of how little compute it draws. Read Comments

Microsoft expands Azure AI with two new models for the Phi-3 family

Microsoft has unveiled a series of updates to its Azure AI platform, including the expansion of the Phi-3 family of small language models (SLMs). The company has added two new models to the family - Phi-3.5-MoE and Phi-3.5-mini - which are designed to enhance efficiency and accuracy. Among the key benefits of Microsoft's new models are their multilingual capabilities - they now support more than 20 models. Phi-3.5-MoE, a 42-billion parameter Mixture of Experts model, combines 16 smaller models into one. By doing this, Microsoft is able to combine the speed and computational efficiency of smaller models with the quality and accuracy of larger ones. Phi-3.5-mini is significantly smaller, at 3.8 billion parameters, however its multilingual capabilities unlock a broader global use case. It supports Arabic, Chinese, Czech, Danish, Dutch, English, Finnish, French, German, Hebrew, Hungarian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish and Ukrainian. Microsoft says that Phi-3.5-mini serves as an important update over the Phi-3-mini model, launched two months ago, based on user feedback. In addition to two new models, Microsoft has also introduced several new tools and services within Azure AI to facilitate easier extraction of insights from unstructured data. More broadly, Microsoft will launch the AI21 Jamba 1.5 Large and Jamba 1.5 models on Azure AI models as a service, offering long context processing abilities. Other announcements included the general availability of the VS Code extension for Azure Machine Learning and the general availability of Conversational PII Detection Service in Azure AI Language. "We continue to invest across the Azure AI stack to bring state of the art innovation to our customers so you can build, deploy, and scale your AI solutions safely and confidently," stated Azure AI Platform Corporate VP Eric Boyd.

OpenAI and Google, get ready: Microsoft is launching its most powerful AI models to date - Softonic

Microsoft has taken a huge step in the field of artificial intelligence with the launch of three new models from the Phi-3.5 series, which not only stand out for their power and versatility, but also for outperforming those of other tech giants, such as OpenAI, Google, and Meta in various benchmark tests. This release includes the Phi-3.5-mini-instruct, Phi-3.5-MoE-instruct, and Phi-3.5-vision-instruct, each one designed to address specific tasks, such as basic reasoning, image processing, and complex data analysis. The Phi-3.5-mini-instruct model, with 3.820 billion parameters, has been optimized for computationally limited environments, offering outstanding performance in tasks such as code generation, solving mathematical problems, and logic-based reasoning. Despite its compact size, this model outperforms larger ones, such as the Llama-3.1-8B-instruct and the Mistral-7B-instruct, in tests like RepoQA, which measures code comprehension in long contexts. On the other hand, the Phi-3.5 MoE (Mixture of Experts) is Microsoft's first model that combines different submodels specialized in various tasks. With 42 billion active parameters and an architecture designed to maintain efficiency, this model stands out in code comprehension, mathematics, and multilingual language, often outperforming larger models in specific tests. Finally, the Phi-3.5 Vision Instruct integrates text and image processing capabilities, making it ideal for complex tasks such as optical character recognition, understanding graphics and tables, and summarizing videos. This multimodal model was trained with a combination of synthetic and public datasets, allowing it to efficiently handle visual tasks with multiple frames. All these models are available on the Hugging Face platform under an MIT license, which allows their use and modification without restrictions, and promotes innovation in both commercial applications and advanced AI research.

Microsoft's Phi-3.5 Models: A Game-Changer in the AI Race?

Given Microsoft's huge strides with the Phi-3.5 models in the artificial intelligence sector, these models embody the future of AI competitiveness. With many cutting-edge features, supreme efficiency, and strong protocols for security, these models define a better momentum for the future of AI competitions. These models can bring about changes to hundreds of sectors by enhancing the capabilities of artificial intelligence, making it more accessible. As the environment of artificial intelligence grows, such AI technologies have the prospect to trade the entirety of the world. 1. What are Microsoft's Phi-3.5 models, and what distinguishes them from other AI models? Microsoft's Phi-3.5 models are a new series of AI models designed to advance performance and efficiency across a variety of applications. The series includes three models: Phi-3.5-mini-instruct, Phi-3.5-MoE-instruct, and Phi-3.5-vision-instruct. Each model is tailored to specific tasks, such as basic reasoning, complex language understanding, and advanced image and video analysis. What sets Phi-3.5 apart is its innovative Mixture of Experts (MoE) architecture in the Phi-3.5-MoE-instruct model, which activates only a subset of parameters at a time, enhancing both efficiency and performance. Additionally, all Phi-3.5 models offer multi-lingual support, covering over 20 languages, and include robust safety measures to ensure ethical AI use. This combination of features allows Phi-3.5 models to deliver high-quality outputs while remaining cost-effective and versatile. 2. How does the Mixture of Experts (MoE) architecture in the Phi-3.5 models work? The Mixture of Experts (MoE) architecture is a key innovation in the Phi-3.5 models, particularly in the Phi-3.5-MoE-instruct model. It combines 16 smaller expert models within a single framework. During training and inference, only a subset of these experts is activated -- specifically, 6.6 billion parameters at a time. This approach allows the MoE model to specialize in different tasks and deliver high-quality outputs with reduced latency and computational cost. By activating only the necessary experts, the MoE model optimizes resource usage and improves efficiency, making it suitable for complex reasoning and language understanding tasks. This architecture not only enhances performance but also ensures that the model remains agile and responsive in various applications. 3. What are the specific applications of the Phi-3.5-mini-instruct model? The Phi-3.5-mini-instruct model is designed for basic reasoning tasks and supports a broad range of languages. Its lightweight and efficient design make it ideal for applications where computational resources are limited. This model is particularly well-suited for use in environments with constrained hardware, such as mobile devices or embedded systems. Typical applications include simple conversational agents, basic language translation, and straightforward data analysis tasks. Due to its efficiency, the Phi-3.5-mini-instruct model can be integrated into various low-power devices and applications, providing effective AI capabilities without requiring significant computational resources. 4. What advantages does the Phi-3.5-MoE-instruct model offer over traditional AI models? The Phi-3.5-MoE-instruct model offers several advantages over traditional AI models due to its innovative Mixture of Experts (MoE) architecture. By activating only a subset of its 16 smaller expert models at a time, it achieves higher efficiency and performance. This approach reduces computational costs and latency while delivering high-quality outputs. The MoE model excels in complex reasoning and language understanding tasks, making it a powerful tool for applications that require advanced language capabilities. Additionally, the MoE architecture supports over 20 languages, enhancing its versatility. Overall, the Phi-3.5-MoE-instruct model represents a significant leap forward in AI technology, combining specialized performance with resource efficiency. 5. How does the Phi-3.5-vision-instruct model enhance image and video analysis? The Phi-3.5-vision-instruct model is designed to advance image and video analysis capabilities. It improves multi-frame image understanding and reasoning, making it particularly useful for applications in computer vision and multimedia processing. This model enhances the ability to analyze and interpret complex visual data, such as identifying objects in a sequence of images or extracting information from videos. Its advanced capabilities support a range of applications, including surveillance, automated content moderation, and advanced image recognition. By leveraging its enhanced visual processing power, the Phi-3.5-vision-instruct model offers significant improvements in the accuracy and efficiency of visual data analysis. 6. What are the multi-lingual capabilities of the Phi-3.5 models? All Phi-3.5 models support over 20 languages, making them highly versatile for global applications. This multi-lingual capability allows the models to be used in diverse linguistic environments, catering to a wide range of users and applications. For example, businesses operating in multiple countries can leverage these models to provide language support across different regions, improving accessibility and user experience. The ability to handle various languages also enhances the models' utility in tasks such as translation, localization, and international customer support. By supporting a broad array of languages, the Phi-3.5 models facilitate global communication and broaden their applicability in international markets. 7. What safety measures are incorporated in the Phi-3.5 models to ensure ethical AI use? Microsoft has implemented several robust safety measures in the Phi-3.5 models to ensure ethical AI use. These measures include supervised fine-tuning and direct preference optimization to produce outputs that are both helpful and harmless. Supervised fine-tuning involves training the models with carefully curated data to improve their accuracy and reliability. Direct preference optimization aligns the models' outputs with user preferences and ethical guidelines, minimizing the risk of harmful or biased responses. By embedding these safety measures, Microsoft aims to set a new standard for responsible AI development, ensuring that the Phi-3.5 models adhere to ethical guidelines and contribute positively to their intended applications. 8. What are the potential benefits of the Phi-3.5 models for small and medium-sized enterprises (SMEs)? The Phi-3.5 models offer several benefits for small and medium-sized enterprises (SMEs), primarily due to their cost-effectiveness and efficiency. The advanced features of the Phi-3.5 models, such as the MoE architecture and multi-lingual support, make them accessible to SMEs that may have limited resources. By leveraging these models, SMEs can integrate sophisticated AI capabilities into their operations without incurring significant costs. This democratization of AI technology allows smaller businesses to enhance their products and services, improve customer interactions, and streamline their processes. Overall, the Phi-3.5 models enable SMEs to leverage cutting-edge AI technology, driving innovation and competitiveness in their respective industries.