AI Research Atlas

Phi-2

Microsoft · 12 December 2023

2.7B model trained on 1.4T tokens of synthetic and filtered web data matches or beats models up to 25x larger on complex benchmarks, per Microsoft.

Scaled Phi to 2.7B in 14 days on 96 A100s; Microsoft reports it beating Mistral-7B and Llama-2 models on several tasks and matching Gemini Nano 2. Initially research-only in the Azure model catalog.

Date
Tuesday, 12 December 2023
Lab
Microsoft
Kind
open-weights
Access
open weights (restricted license)

Figures

MeasureValueMeasured by
Training1.4T tokens, 14 days on 96 A100s
2.7B parameters		company

Initially offered in the Azure AI Studio catalog for research; licence terms not stated in the blog post.

Sources

  1. www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/

This record was checked against its sources on 6 October 2026. How we check

Related