Phi-3.5-mini-instruct

Phi-3.5-mini-instruct is a lightweight 3.8B-parameter language model from the Phi family (Microsoft), instruction-tuned. It uses a dense decoder-only Transformer architecture and offers a 128K-token context window, targeting memory/compute-constrained environments and latency-bound scenarios.

The model was trained on 3.4 trillion tokens — synthetic data and filtered public websites — with an emphasis on high-quality reasoning content. Post-training included supervised fine-tuning (SFT), proximal policy optimization (PPO), and direct preference optimization (DPO). It supports 22 languages (including Arabic, Chinese, French, German, Japanese, Korean, Spanish).

Phi-3.5-mini achieves competitive results: 69% on MMLU and 86.2% on GSM8K (mathematical reasoning), with strong code (mainly Python), math, and logic capabilities. Released under the MIT license; available via Hugging Face, Azure AI Foundry, and inference providers (vLLM, SGLang).