Phi-4

Released by Microsoft Research in December 2024, Phi-4 packs 14 billion parameters into a dense decoder-only transformer with a 16K-token context window. Training covered 9.8 trillion tokens of filtered web pages, books and Q&A collections Microsoft obtained, and a large share of synthetic textbook-style material written to teach math, logic and coding; safety and instruction tuning used SFT followed by rounds of DPO. Microsoft's simple-evals results put it close to far larger models on MATH and GPQA, yet very low on the SimpleQA factual recall test, so it reasons better than it remembers. Prompts use a chat template with im_start style tags and the model loads through Transformers; official GGUF weights exist, and Ollama carries phi4. A 4-bit quantized copy fits on a 12 GB consumer card, while unquantized bf16 weights need close to 30 GB. Later relatives include Phi-4-mini, Phi-4-multimodal and Phi-4-reasoning variants. All use the MIT license, which permits commercial use and fine-tuning. Best for developers who want a local model for math, logic or structured problem solving on one workstation GPU, mainly in English.