Microsoft Phi

Phi-1 was a 1.3-billion-parameter code model trained on selected web data plus synthetic textbooks and exercises generated with GPT-3.5, according to its paper. Microsoft later described Phi-2 as a 2.7-billion-parameter model in a research blog and model card. The Phi-3 technical report describes Phi-3-mini at 3.8 billion parameters, including context variants, and introduces Phi-3.5-mini, Phi-3.5-MoE, and Phi-3.5-Vision. Phi-3.5-Vision accepts image and text prompts; the MoE variant uses a mixture-of-experts architecture. Performance, licensing, and deployment fit depend on the exact version and evaluation setup; small parameter count alone does not guarantee that a model runs efficiently on every laptop or phone. Phi-1’s “Textbooks Are All You Need” title refers to a specific coding-model experiment that compared curated and synthetic data recipes; it does not establish a universal data-quality-over-scale law. Compact Phi checkpoints may be deployable locally when hardware and runtime support them, but size alone guarantees neither privacy nor latency or quality. The Phi-3-mini materials list 4K and 128K context variants. Phi-3.5-Vision supports single-image and multi-image prompts, while Phi-3.5-MoE uses sparse routing that activates a subset of experts for each token. These facts are specific to the cited checkpoints.