News
Microsoft Launches Phi 3.5-Mini Models
Microsoft announced that its Phi-3.5-Mini-Instruct model, the latest update to its Phi-3 model family, is now available.
The Phi family is Microsoft's assorted compact micro models that can run on a smartphone and operate without an Internet connection. The previous inaugural version, Phi 3, was launched earlier in April, and Microsoft said it's been working on improvements since the release.
"The model underwent a rigorous enhancement process, incorporating both supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures," wrote Microsoft.
The newly introduced Phi 3.5 models feature the 3.82 billion parameter Phi-3.5-Mini-Instruct for basic and fast reasoning, the 41.9 billion parameter Phi-3.5-MoE-instruct for advanced reasoning tasks and the 4.15 billion parameter Phi-3.5-Vision-Instruct, specifically tailored for vision tasks like image and video analysis.
As with the previous version, the Phi-3.5-Mini-Instruct model has a 128K token context length, making it suitable for tasks requiring extensive context, such as long document summarization and information retrieval. It was trained on 3.4 trillion tokens using 512 H100-80G GPUs over 10 days, while the Vision Instruct model underwent training on 500 billion tokens with 256 A100-80G GPUs over a span of six days.
According to Microsoft's own testing (found here), all three models deliver strong performance on several third-party benchmark tests, outperforming other AI providers such as Google's Gemini 1.5 Flash, Meta's Llama 3.1 and OpenAI's GPT-4o (in certain instances).
While Microsoft's benchmarks found that it did outperform competitors in some situations, due to the smaller nature of the models, they aren't equipped to go head-to-head on every task other large language models can take on.
"The model simply does not have the capacity to store too much factual knowledge, therefore, users may experience factual incorrectness," said Microsoft. "However, we believe such weakness can be resolved by augmenting Phi-3.5 with a search engine, particularly when using the model under RAG settings."
Microsoft has included the Phi 3.5 Mini-Instruct in some of its own services and is available in Azure AI Studio, making it accessible to a wide range of users.
All three models are open sourced and available freely through the MIT license. The license does stipulate that the software is provided "as-is," and the company said cannot be held liable for any possible damages.
Microsoft said it will continue to take feedback from the community to refine and improve the Phi-3 model family.