Microsoft Launches New Phi-3.5 Models, Outperforms Google and More
Microsoft has recently unveiled its latest Phi-3.5 models, which include Phi-3.5-MoE-instruct, Phi-3.5-mini-instruct, and Phi-3.5-vision-instruct. These models are designed for various reasoning and vision tasks, with parameters ranging from 3.8 billion to 41.9 billion.
Phi-3.5-MoE-instruct
The Phi-3.5-MoE-instruct model boasts an impressive 42 billion parameters and excels in handling advanced reasoning tasks. It has demonstrated significant improvements in reasoning capabilities, surpassing models like Llama 3.1 8B and Gemini 2 9B across different benchmarks.
Phi-3.5-mini-instruct
On the other hand, Phi-3.5-mini-instruct, with 3.8 billion parameters, is tailored for basic reasoning tasks and offers quick responses. Despite its lightweight design, it outperforms larger models like Llama 3.1 8B and Mistral 7B, making it a competitive option in various applications.
Phi-3.5-vision-instruct
Phi-3.5-vision-instruct, equipped with 4.2 billion parameters, is focused on visual tasks such as image and video analysis. It showcases superior performance in tasks like multi-frame image understanding and reasoning, outperforming models like OpenAI GPT-4o on certain benchmarks.
The Phi-3.5 models have been trained using advanced technologies and methodologies, including supervised fine-tuning, proximal policy optimization, and direct preference optimization. These models support multilingual applications and offer extended context lengths, making them versatile for a wide range of tasks.
Key Features and Use Cases
The Phi-3.5 models are developed to be used in memory and compute-constrained environments, catering to applications that require strong reasoning in areas like code, mathematics, and logic. They are also ideal for generative AI-powered features and are suitable for both commercial and research purposes.
Phi-3.5 models are now available on the AI platform Hugging Face under an MIT license, ensuring accessibility for various applications. This release is in line with Microsoft's commitment to providing efficient and versatile open-source AI tools for the community.
For more information on the Phi-3.5 models, you can visit the official Hugging Face website.
