ModelsMicrosoft
Microsoft named Aion Instruct, a 2.3B on-device model, as the replacement for Phi Silica in Windows
32Ktoken context, against 3.5K for Phi Silica
Microsoft set out on 13 November how Aion Instruct will replace Phi Silica as Windows' on-device language model. Aion Instruct has 2.3 billion parameters, against 3.8 billion for Phi Silica. Its context limit is 32K tokens, up from 3.5K. It reaches Windows Insider devices from 20 November, with retail devices planned for January 2027. It is also coming to supported GPUs, including NVIDIA RTX and AMD RDNA4 cards. The dates are Microsoft's plans.
Microsoftfrom the source itself