Phi-3

Microsoft's Phi-3 family, first published in April 2024 with a June 2024 refresh of the Mini model, consists of dense decoder-only models: Mini at 3.8B parameters, Small at 7B and Medium at 14B, each offered with a short context (4K or 8K tokens) and a 128K version, plus a 4.2B vision model built on Mini. The bet behind the family was that heavily filtered web text combined with synthetic, textbook-style material buys reasoning ability per parameter; Mini saw 4.9 trillion training tokens and was then aligned with supervised fine-tuning and DPO. Weights are on Hugging Face for Transformers, with official ONNX builds for CPU, CUDA and DirectML and GGUF files for llama.cpp, and Ollama carries phi3 tags. A quantized Mini runs comfortably on an ordinary laptop, while Medium needs a reasonably capable GPU even at 4-bit. The MIT license allows commercial use without extra terms. Expect English-first behavior, thin factual recall, and code skills limited mostly to common Python packages. Later Phi-3.5 and Phi-4 releases beat it, so Phi-3 now makes most sense for constrained edge deployments and existing pipelines already tuned around it.