Microsoft Phi-4 models: small enough to run on the device

Microsoft contribution here is the Phi family: compact models tuned to punch above their parameter count. They are open weight and small enough for edge and on-device deployment, which is the point of them.

Phi is not competing with frontier models on hard reasoning. It exists for cases where latency, privacy or hardware constraints rule out a hosted call — classification on device, offline assistants, embedded features.

• Benchmark against a comparable model from a larger vendor before committing production traffic — line-ups this size get less independent testing.
• Check the context window against your longest input — the largest here is Phi-4 Mini Instruct at 131.1K tokens.
• Request current rates directly: 3 of the listed models publish no public per-token price.
• Confirm support terms and deprecation policy, which smaller vendors document less thoroughly.