Phi-4

Microsoft's Phi-4 is a 14-billion-parameter dense decoder-only Transformer trained on ~9.8 trillion tokens of high-quality synthetic and curated organic data. It became a flagship example of the "small model, big reasoning" trend: strong STEM, math, and code performance from a model that runs on a single workstation GPU. Weights are fully open under the MIT license (verified on the Hugging Face repo), and it is supported by vLLM, llama.cpp, Ollama (ollama pull phi4), and Azure AI Foundry. Note: Phi-4 itself is text-only — multimodal variants (Phi-4-multimodal, Phi-4-mini) are separate releases.

Provider-reported (Phi-4 Technical Report, arXiv 2412.08905, simple-evals framework): MMLU 84.8, GPQA 56.1, MATH 80.4, MGSM 80.6, HumanEval 82.6, HumanEval+ 82.8, ArenaHard 75.4 — beating GPT-4o-mini and Llama-3.3-70B on MATH, GPQA and HumanEval. Independent: third-party benchmark runs beyond provider tables are thin; coverage is press-level (independent tech outlets verified the launch specs and HF release, but did not reproduce evals). Treat head-to-head wins as provider-claimed.