Phi-4 Mini Instruct
Phi-4 Mini Instruct combines a 3.8B text model with a 128K context and a documented tool-call format. Its compact size and long input limit make it a different deployment choice from the 14B Phi-4 release.
What distinguishes this release
The architecture uses grouped-query attention, shared input/output embeddings and a 200K vocabulary. The model card documents tool definitions inside dedicated tool tokens and identifies factual knowledge capacity as a limitation.Phi-4 Mini Instruct official model card
Build the task around its input and output
Practical guidance. For a small local assistant, separate the language task from access to current facts. Let retrieval supply the records and let the model select or explain them. Include long inputs only when they contain information that the answer actually needs.
Review the result in its intended setting
Practical guidance. Treat a proposed tool call as data to validate. Check its tool name, arguments and allowed scope before execution. Review multilingual outputs in the intended language; a larger vocabulary is not evidence that every language performs equally well.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | microsoft/Phi-4-mini-instruct |
| Context window | 131,072 tokens Published checkpoint context; serving engines may impose lower limits Phi-4 Mini Instruct official model card |
| Model size | 3.8B densePhi-4 Mini Instruct official model card |
| Input and output | Text input and text outputPhi-4 Mini Instruct official model card |
| Weight license | MITPhi-4 Mini Instruct official model card |