微软与英伟达推出搭载RTX Spark的Surface Laptop Ultra

At an event in San Francisco, Microsoft and Nvidia teamed up to show off the RTX Spark chip, along with the Surface Laptop Ultra and new features in Windows 11 that use the new system-on-a-chip's capabilities. The device, along with the Surface RTX Spark Dev Box and the broader Spark ecosystem is available for pre-order today, with devices shipping on October 16.

Microsoft CEO Satya Nadella, Nvidia CEO Jensen Huang, and Windows and Devices head Pavan Davuluri took turns on stage.

"We've reinvented the personal computer completely," Huang said.

Nadella described the event as one that would be building the "frontier ecosystem," anchoring it in Agent 365, Microsoft IQ, and Copilot. Davuluri describes Windows 11 as being the most secure platform for AI agents, highlighting containment for agents.

Windows 11 and Surface Laptop Ultra

The Surface Laptop Ultra will start at $2,599, while the Surface RTX Spark Dev Box will begin at $5,999. Both are available for pre-order. They, along with PCs from Asus, Dell, Lenovo, HP, and MSI will ship next Friday, Oct. 16.

Davuluri showed the Surface Laptop Ultra running Github Copilot and playing Gears of War: E-Day at 60 frames per second. He said that "hundreds" of AAA titles and indies will be available day one, and that Call of Duty will come to RTX Spark in 2027.

The 4.5-pound system has the brightest peak HDR on a laptop, he said. The system also features a magnetic USB-C charger it calls "magnetic connect." The connector plugs in, similar to Apple's MagSafe or the old Surface Connect ports, except using a standardized port.

An impressive video demo using multiple local and cloud models across Blender, Photoshop, ComfyUI, and Unreal Engine to make a workload that took hours instead of days was impressive, but also highly produced.

Microsoft compared the Spark to the M5 Pro in a MacBook Pro, promising performance up to 2.1 times faster to first token, 4.3 times faster image generation, and 6.2 times faster video generation.

On Windows 11, there will be new quick actions in the Search menu in the Task Bar, making quick changes to settings, taking screenshots, or arranging windows using natural language. The Taskbar will also launch Copilot if the request calls for it, and you can jump from Search to Copilot.

Davuluri also described local models running from other vendors, including Hermes, OpenClaw, and Perplexity, performing a demo with Perplexity Computer. He also stated that Meta's Muse will come to Windows. There will be a new setup for Windows AI agents to make them easier to configure.

Davaluri also announced DGX Spark and Station for Windows, a "supercomputer" to bring "last quarter's frontier to your desktop." He showed off desktops from HP and Dell that will come "later this year." Up to this point, both have been exclusive for Linux.

Copilot and Local Agents

Microsoft's Jacob Andreou detailed the recent launch of the new Copilot, which is now getting access to local files, local models, and the ability to take actions on account of the user. It will still use cloud models, but will switch to local models when privacy is called for.

In one example, Andreou was grabbing tax data for his accountant. The system used a mix of local and cloud models to search for documents, rename them, and zip them into a folder. Copilot also wrote an email to the accountant to send on his behalf, which he approved. These features, he said, are coming over the next few months.

Containers, Trust, and Hybrid Ingelligence

Daviluri focused on the technical details of agents, including Microsoft Execution Containers, which is moving to general availability. These are sandboxes that are integrated with the OS to decide what data is shared with agents.

Windows will also require agents to attribute their actions to themselves. There will be IT actions for enterprises that want to set agentic rules once and use them across their fleet of devices.

Microsoft's Divya Venkataramu demonstrated this system, showing how she had access to documents even when agents didn't.

Davuluri also described hybrid intelligence, a mix of agents running locally and in the cloud to reduce token budgets. Microsoft's HydraFusion, which balances local and cloud agents, is coming from GitHub to Windows on October 15 for GitHub Copilot, the CLI, and VS Code, and will use new models like MAI Code 1.1 Flash and Nvidia's Nemotron. Using advanced quantization, DeepSeek V4 Flash can be used at 1.6 bits to run locally at 60GB of memory, he said.

Windows ML will serve as the runtime to deploy models across GPU, NPU, and CPU. Llama.cpp will come to Windows ML to give developers chances to experiment with new models at launch.

Microsoft's Kayla Cinnamon demonstrated Hybrid intelligence on a Surface RTX Dev Box by prompting GitHub Copilot to detect unused codes, packages, and feature flags to clean up code. With an Efficiency mode, Copilot selects a model automatically based on how complex a task is. A more complex prompt used GPT-6 Luna, while separate sub-agents ran locally.