Phi-2

Phi-2 is Microsoft's 2.7-billion-parameter small language model, released in December 2023 and trained on 1.4 trillion tokens of 'textbook-quality' real and synthetic data over 14 days on 96 A100 GPUs. Microsoft reported it matching or beating models up to 25x its size — including Llama 2 7B, Mistral 7B, and Google's Gemini Nano — on reasoning benchmarks, making it the defining demonstration of the data-quality-over-scale thesis.

License history worth knowing: Phi-2 launched under a Microsoft Research License restricted to non-commercial research use; it was later re-released under MIT, which the official repository now uses. Even under MIT, it predates modern safety alignment, so it can produce inaccurate or biased content and should not be deployed as a user-facing assistant without guardrails.