InfoWorld

Machine Learning News

Sam Witteveen

<h1>Microsoft Unveils Advanced Phi 3.5 Models Surpassing Google and Meta</h1>
<p>Microsoft has recently introduced a suite of powerful new language models under the Phi 3.5 series, named Phi 3.5-mini-instruct, Phi 3.5-MoE-instruct, and Phi 3.5-vision-instruct. These models have shown exceptional performance, outperforming similar offerings from Meta and Google in several benchmarks. Available on Hugging Face under the MIT License, these models allow for extensive flexibility in usage and customization.</p>
<h2>Key Features and Models</h2>
<ol>
<li>
<p><strong>Phi 3.5-MoE-instruct</strong>: This Mixture of Experts (MoE) model comes with 42 billion parameters and dynamically engages different subsets of experts, thereby improving performance on a variety of benchmarks.</p>
</li>
<li>
<p><strong>Phi 3.5-mini-instruct</strong>: A more efficient model with 3.8 billion parameters, designed to handle long context tasks and showing significant improvements in multimodal tasks including languages like Arabic and Chinese.</p>
</li>
<li>
<p><strong>Phi 3.5-vision-instruct</strong>: Integrating both visual and textual data, this model is ideal for complex applications such as medical imaging and autonomous vehicles.</p>
</li>
</ol>
<h2>Training and Performance</h2>
<p>These models have been trained on extensive datasets using Nvidia GPUs, bestowing them with advanced reasoning abilities and robust multilingual support. However, they do fall short in some aspects when compared to OpenAI’s GPT-4o-mini.</p>
<h2>Applications and Flexibility</h2>
<p>The Phi 3.5 models cater to a broad range of applications, from reasoning and code generation to image and video analysis. Their release under the MIT License ensures that developers can easily download, customize, and deploy these models for various commercial and non-commercial purposes.</p>
<p>With the Phi 3.5 series, Microsoft has significantly bolstered its AI capabilities, offering advanced tools for industries like healthcare, autonomous driving, and beyond, making these models a competitive choice against contemporary AI from top tech behemoths like Google and Meta.</p>


Collections

Microsoft has launched new Phi 3.5 language models that outperform Google and Meta in several benchmarks. The models include Phi 3.5-mini-instruct, Phi 3.5-MoE-instruct, and Phi 3.5-vision-instruct, each designed for specialized tasks ranging from multimodal processing to handling extensive datasets. Available under the MIT License on Hugging Face, they offer significant flexibility for developers in various applications including healthcare and autonomous driving.

Microsoft’s New Phi 3.5 LLM Models: Surpassing Meta, Google, and OpenAI