Microsoft's Phi-3 AI Models Outperform Competitors, Excel in Benchmarks

Microsoft has announced the release of its new language models under the Phi-3 series, featuring advancements in AI capabilities. The Phi-3-mini, a 3.8 billion parameter model trained on 3.3 trillion tokens, is designed to rival major models like Mixtral 8x7B and GPT-3.5. Additionally, the Phi-3-medium, with 14 billion parameters trained on 4.8 trillion tokens, achieves a 78% score on the MMLU benchmark and 8.9 on MT-bench. The Phi-3 14B model outperforms the Llama-3 8B, GPT-3.5, and Mixtral 8x7b MoE in most benchmarks, while the Phi-3 mini also surpasses the Llama-3 8B in MMLU and HellaSwag. The series also includes the Phi-3 7B model, which surpasses the Llama-3 7B model in performance, scoring 75.3 on MMLU. These models are part of Microsoft's effort to enhance the open-source community, sharing similar architecture with Llama-2.