This open-source model was created by Microsoft. You can find the release blog post here. The model is available on the huggingface hub: https://huggingface.co/microsoft/Phi-3-mini-128k-instruct. The model has 16x3.8B parameters with 6.6B active parameters, and supports up to 128K token contexts. Even though this model supports system messages, we evaluate this model as user-message-only model (the persona is induced by sending the user message "You are <persona>" followed by a manually set "OK" as the assistant's response) as it worked better.