Microsoft Expands Compact AI Lineup With Phi 3.5 Models
Microsoft releases three new compact AI models under the Phi 3.5 family, designed for on-device tasks like fast reasoning and image analysis. The mini-instruct model outperforms several larger competitors in specific benchmarks despite its small size.
Microsoft unveils its new Phi-3.5-Mini-Instruct model alongside two other compact AI models, expanding its lineup of lightweight systems that run locally on smartphones without an internet connection. The updated models undergo rigorous enhancement processes, including supervised fine-tuning and direct preference optimization, to ensure precise instruction adherence and robust safety measures.
The new Phi 3.5 family includes the 3.82 billion parameter Mini-Instruct for fast reasoning, a 41.9 billion parameter MoE-instruct for complex reasoning, and a 4.15 billion parameter Vision-Instruct for image and video analysis. The Mini-Instruct model maintains a 128K token context length, making it highly effective for long document summarization and extensive information retrieval tasks.
Microsoft trains the Mini-Instruct model on 3.4 trillion tokens using 512 H100-80G GPUs over 10 days, achieving strong results that outperform competitors like Google's Gemini 1.5 Flash and Meta's Llama 3.1 in specific benchmark tests. However, Microsoft acknowledges that the smaller model lacks the capacity to store vast amounts of factual knowledge, meaning users experience factual incorrectness in certain scenarios.