Microsoft is planning to introduce “NPU-friendly” versions of the DeepSeek-R1 AI model for Copilot+ computers, starting with Snapdragon X devices, then moving on to Intel Lunar Lake and AMD Ryzen AI 9 processors. The initial release will be called DeepSeek-R1-Distill-Qwen-1.5B and can be accessed through the Microsoft AI Toolkit by developers. Later, they’ll also offer 7B and 14B versions.
According to a Microsoft blog post announcing DeepSeek R1 support, these optimized models enable developers to construct and launch AI-driven applications that run smoothly on devices, leveraging the potent NPUs in Copilot+ PCs. Our Phi Silica project maximizes efficient inference, ensuring swift first token times and impressive throughput rates without significantly draining battery life or using up excessive PC resources. Moreover, we utilize Windows Copilot Runtime (WCR) to adapt across the wide range of devices within the Windows ecosystem with ONNX QDQ format.
In the blog post, Microsoft explains their method for enabling R1 models to run on devices with NPU (Neural Processing Unit) hardware. This is achieved by implementing a sliding window design that provides quick access to the first token and support for long contexts, even though the hardware stack doesn’t offer dynamic tensor support. Furthermore, they employ a 4-bit QuaRot quantization method to maximize the benefits of low-bit processing.
Microsoft has announced that the 1.5B Distilled R1 model will soon be available for use, and it can be accessed through the AI Toolkit extension in Visual Studio Code. Developers can explore and test DeepSeek R1 on compatible Copilot+ systems using the Playground feature locally. Furthermore, Microsoft is providing these AI models not only for local use but also in the cloud via Azure AI Foundry. The Azure AI Foundry makes DeepSeek R1 available on a secure, scalable enterprise platform, allowing businesses to effortlessly incorporate sophisticated AI while adhering to service level agreements, security standards, and responsible AI principles, all with Microsoft’s renowned reliability and innovation as backing.
Microsoft swiftly jumped on board with DeepSeek R1, while U.S. tech companies express concern over its emergence. OpenAI alleges that DeepSeek illegally used proprietary code in creating their AI model, a project that reportedly required less than $10 million to complete. This is significantly lower compared to the enormous financial investment made by U.S. firms in AI development, which currently totals billions of dollars.
Read More
- EigenLayer restaking protocol’s slashing testnet now live
- FIL PREDICTION. FIL cryptocurrency
- 3 Spider-Man Villains That Should Appear in the Next PlayStation Game
- POL PREDICTION. POL cryptocurrency
- Marvel Rivals Season 1 is bringing the Fantastic Four — here’s the Invisible Woman in action
- Grantchester season 9 cast: Meet the characters in ITV drama
- EastEnders airs huge Cindy Beale update in early iPlayer edition
- Whales and retail traders keep Arbitrum from falling — Here’s how
- Looks Like DOOM 64 Is Getting a Native PS5 Port
- Coronation Street’s Colson Smith breaks silence on when he was told about Craig Tinker’s axing
2025-01-30 02:39