Phi-3 Mini has two context windows and several deployment paths

Microsoft has released Phi-3 Mini, a 3.8-billion-parameter language model, through Azure AI Studio, Hugging Face and Ollama. Its April 23 announcement includes both 4K and 128K context versions. The larger Phi-3 Small and Medium models remain promised for the coming weeks, even though Microsoft’s launch material already presents results for them. Microsoft announcement

For developers, there are three separate decisions inside that launch: which context variant to obtain, which runtime can execute it, and what a particular device can actually sustain.

File photo: Microsoft buildings, efes, 2015. CC0 1.0, via Wikimedia Commons. Recompressed and responsively cropped; not the Phi-3 launch.

Two launch variants, different configurations

The original Hugging Face repositories contain both Mini variants in April 22 file-upload commits. The 4K configuration sets its maximum position count to 4,096. The 128K version sets 131,072 and supplies additional position-scaling factors, while retaining 4,096 as its original context setting. These are distinct model packages, rather than a promise that changing a prompt-length setting turns the shorter version into the longer one. 4K configuration, 128K configuration

That distinction matters when interpreting Microsoft’s phone demonstration. The original technical report describes a four-bit-quantized Mini running fully offline on an iPhone with an A16 Bionic chip, generating more than 12 tokens a second. It doesn’t establish that a full 128K conversation runs at that speed on that phone. The demonstrated device, quantization and workload aren’t a universal hardware specification. Original technical report

Runtime support has its own timetable

NVIDIA’s announcement offers a 128K Mini trial through its NIM interface. In the same article, it says the TensorRT-LLM example implementations will become available soon. Access to a hosted trial therefore doesn’t establish that the corresponding local example is already released. NVIDIA announcement

The launch model cards also describe different execution paths. Their sample loads repository-provided code through Transformers; the hardware section directs older NVIDIA GPUs toward eager attention and points to separately optimized ONNX models for other targets. The 4K card additionally links a quantized GGUF option for CPU use. Those paths shouldn’t be collapsed into one claim that every download works identically on every laptop. Launch model card

Both model repositories carry the MIT licence, including its notice-preservation requirement and warranty disclaimer. That’s a distribution permission, not evidence that an application has passed its own quality checks. Launch licence

The cards describe intended English-language commercial and research use and an October 2023 data cutoff. Microsoft’s report separately acknowledges limited factual knowledge and reports benchmark results from its own evaluation pipeline. A longer context window gives supplied material more room; it doesn’t update the model’s stored knowledge or turn vendor testing into an independent assessment. Launch card, Technical report