Phi-3-vision-128k-instruct

Phi-3-vision-128k-instruct is Microsoft's lightweight open multimodal model, released May 21, 2024. With 4.2 billion parameters, it combines an image encoder, connector, projector, and the Phi-3 Mini language model, trained on 500 billion vision and text tokens with 512 H100 GPUs over 1.5 days. It supports 128K tokens of context.

The model targets memory-constrained and latency-bound scenarios: general-purpose AI features on phones and edge devices that need strong reasoning, especially math and logic. It underwent supervised fine-tuning and direct preference optimization for instruction adherence. Weights ship under the MIT license with an open release type.