phi-3-mini-4k-instruct-june-2024

phi-3-mini-4k-instruct-june-2024 is a 3.8 billion-parameter lightweight language model developed by Microsoft. It is a refined version of the original Phi-3-mini released in April 2024, specifically updated in June 2024 to enhance instruction following, reasoning, and structured output capabilities. The model belongs to the Phi-3 family of small language models (SLMs) and is designed to deliver performance comparable to significantly larger models while remaining efficient enough to run on consumer hardware.

Built on a transformer decoder-only architecture, the model was trained on 4.9 trillion tokens consisting of high-quality synthetic data and filtered public web content. The June 2024 update specifically improved the model's ability to handle multi-turn conversations and introduced explicit support for the <|system|> tag. It also showed substantial gains in benchmarks for JSON and XML structured outputs.

The model is optimized for reasoning-dense tasks such as mathematics, coding, and logical analysis. It supports a context window of 4,000 tokens (4,096), balancing low latency and memory efficiency for edge and mobile deployment scenarios. The model was fine-tuned using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to align with human preferences and safety guidelines.