Definition
Plain language
Microsoft's small open language model in the 14-billion-parameter range.
As stated in the literature
A Microsoft Research open-weight foundation model around 14B parameters, used in mechanistic interpretability work on sycophancy circuits.
Why it matters: A capable but tractable open model makes mechanistic interpretability studies feasible at a useful scale.
For example, researchers might trace sycophancy-related attention heads inside Phi-4 to understand how flattery is implemented.
Heard on the show
“Three instances of a large language model — they used GPT-4o-mini, DeepSeek-V3, and Phi-4 in different runs — take turns talking in a shared room.”Episode 073 — When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving