Microsoft’s Phi-3.5-MoE Competes with Gemini 1.5 Flash, Now Accessible in Azure AI Studio and GitHub

Last month, Microsoft unveiled the cutting-edge Phi-3.5 family of lightweight AI models, which come with an array of enhancements. The standout among these is the Phi-3.5-MoE, marking the first model in the Phi series to incorporate Mixture of Experts (MoE) technology.

Microsoft has now announced that the Phi-3.5-MoE model is readily available in Azure AI Studio and GitHub via a serverless API. This feature allows developers to seamlessly integrate the Phi-3.5-MoE model into their workflows and applications without the need to manage any underlying infrastructure.

The Phi-3.5-MoE model, alongside other Phi-3.5 models, can be accessed in several regions, including East US 2, East US, North Central US, South Central US, West US 3, West US, and Sweden Central. As a serverless offering, developers benefit from a pay-per-use pricing structure, which is set at $0.00013 per 1,000 input tokens and $0.00052 per 1,000 output tokens.

In various AI benchmarks, the Phi-3.5-MoE has demonstrated superior performance relative to nearly all other open models in its category, such as Llama-3.1 8B, Gemma-2-9B, and Mistral-Nemo-12B, notably utilizing fewer active parameters. Microsoft asserts that its performance rivals, if not slightly surpasses, Google’s Gemini-1.5-Flash, one of the leading closed-source models in this domain.

The MoE model features a total of 42 billion parameters, of which only 6.6 billion are activated, supported by 16 experts. The team at Microsoft Research designed this model from the ground up to enhance performance, increase multilingual capabilities, and reinforce safety protocols. Additionally, rather than relying on conventional training techniques, the Microsoft Phi team has pioneered a novel training method known as GRIN (GRadient INformed) MoE. This approach has led to significantly improved parameter utilization and expert specialization, achieving markedly higher quality outcomes compared to traditional training modalities.

With its exceptional performance metrics and accessibility, the Phi-3.5-MoE is set to empower developers and drive innovation within the AI ecosystem. Its serverless model and consumption-based pricing are further dismantling barriers to entry, enabling more developers to access advanced AI capabilities than ever before.

Source: Microsoft

Source