Deepsek conquered the world of mobile and it is now expanding into Windows – surprisingly with full support of Microsoft. Yesterday, the software giant added the Depusic R1 model to its Azure AI Foundry, so that the developers would be allowed to test and create cloud-based apps and services with it. Today, Microsoft announced that it was bringing the distilled versions of R1 into Copilot+ PC.
The distilled model will first be available for devices powered by Snapdragon X chips, with Intel Core Ultra 200th 200th Processor and then with AMD Rvenie AI9 based PC.
The first model will be the Dipsec-R1-Dystil-Quven-1.5B (ie 1.5 billion parameter model), with large and more competent 7B and 14B models coming soon. These will be available for download from Microsoft’s AI toolkit.
These models had to be tickled to customize Microsoft on devices with NPU. Operations that depend a lot on the memory access on the CPU, while the transformer blocks such as computational-lonity operations run on the NPU. With adaptation, Microsoft managed to achieve rapid time for the first tokens (130ms) and rate a 16 tokens per second for small signals (under 64 tokens). Note that a “token” is similar to a tone (significantly, a token is usually more than one character tall).
Microsoft Openai (Cutgpt and GPT-4o manufacturers) have been invested in a strong supporter and depth, but it seems that it does not play a favorite-GPT model (Openai), Llama (Meta), Mistral in its azure playground Is. An AI company), now Deepsek too.
Deepsek R1 at Azure AI Foundry Playground
Anyway, if you are more in the local AI, then download the AI Toolkit for the first VS code. From there, you should be able to download the model locally (eg “Deepsek_r1_1_5” is a 1.5B model). Finally, try in the playground and see how smart this distilled version of R1 is.
“Model distillation”, sometimes called “knowledge distillation”, is a large AI model (671 billion parameters in full deepsek R1) and for a small model (eg 1.5 billion parameters) to your knowledge. To move as much as possible. , This is not an ideal process and the distilled model is less capable than the full model – but its small size allows it to run directly on consumer hardware (cost thousands of dollars instead of dedicated AI hardware).
