Public transcript

Microsoft Windows Event with Satya Nadella and Jensen Huang | LIVE

By: Yahoo Finance
Length: 1 h 0 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This Microsoft Windows event, featuring Satya Nadella and Jensen Huang, announces hybrid intelligence on Windows: AI models, runtimes, and silicon working together across cloud and local devices. Key announcements include GitHub Copilot hybrid routing, new Windows search actions, Copilot with local context, actions, and models, MXC agent security, and the Surface Laptop Ultra built on Nvidia RTX Spark. Nadella and Huang close with a fireside chat on their decades-long partnership, agent security, and the reinvention of the personal computer.

Key points

Microsoft announced hybrid intelligence on Windows, routing tasks between cloud and local models via GitHub Hydro Fusion.
MAI code 1.1 Flash was quantized to 3 bits, reducing size nearly 80% while preserving coding quality, with a 256K context window on device.
DeepSeek Flash V4, a 284 billion parameter model, runs locally at 1.66 bits in about 60 GB and outperforms GPT5 on encoding and reasoning benchmarks.
Hybrid intelligence comes to the GitHub Copilot app, CLI, and VS Code starting October 15th.
The new Copilot gains local context, local actions, and local models, letting it ground over PC files and act with permission.
Surface Laptop Ultra, built around Nvidia RTX Spark, offers up to 128 GB unified memory and up to a petaflop of AI compute.
MXC was presented as a new operating system primitive for agent security, with containment, identity, observability, and governance.
Jensen Huang said RTX Spark took over 4,000 engineering years and reinvented the personal computer, running DirectX, OpenGL, and CUDA natively.

Questions it answers

00:00Hybrid intelligence and local models

What is hybrid intelligence and which models can now run locally on Windows?

Hybrid intelligence combines cloud and local models, runtimes, and silicon. Microsoft has quantized MAI code 1.1 Flash to 3 bits, cutting its size by nearly 80%, and Nvidia's new Neatron model (over 70 billion parameters) quantizes to 2-bit using just over 20 GB, ideal for RTX Spark.

MAI code 1.1 Flash quantized to 3 bits, shrinking size nearly 80% while preserving coding quality.
DeepSeek Flash V4, a 284 billion parameter model, runs locally in about 60 GB at 1.66 bits.
Llama CPP is coming to Windows ML, giving developers open source models on day one.

03:00GitHub Copilot demo

How does hybrid intelligence work in GitHub Copilot on Windows?

Kayla demonstrates that Copilot's auto mode routes simple tasks to local models like MAI code 1.1 Flash, which keep working even in airplane mode, while complex tasks use cloud models like GPT6 Luna. Local sub sessions generated 1.66 million tokens for free, and automations like issue triage can run overnight on local models without spending credits.

Auto efficiency in the model picker selects a local model for simpler tasks.
Local sub sessions ran 1.66 million tokens sent to the model for free.
Automations can run on a cadence or trigger, explicitly using the local model.

07:46Windows search and Copilot

What's new in Windows search and the Copilot app?

Starting this fall, Windows 11 taskbar search supports thousands of quick actions like dark mode, do not disturb, screenshots, and sending texts. The new Copilot app gains local context (grounding over files on your PC), local actions (moving files, changing settings with permission), and local models for when cost or privacy matter, as shown in an Autopilot tax filing demo.

Hybrid intelligence in GitHub Copilot app, CLI, and VS Code arrives October 15th.
Copilot gets local context, local actions, and local models.
Autopilot organized tax files locally and zipped them without sending details to the cloud.

16:00Ecosystem and devices

What devices and ecosystem partners were announced?

Microsoft announced a new get started experience for running agents on mini desktop PCs, Copilot Plus PCs handling over two trillion inferences per month locally, and Surface Laptop Ultra built around Nvidia RTX Spark with up to 128 GB unified memory and a petaflop of AI compute. RTX Spark laptops from Lenovo, Asus, Dell, MSI, HP, and Surface are available for pre-order, plus DGX Station for Windows running frontier-class models locally.

Over 40% of laptops built for business are Copilot Plus PCs.
Surface Laptop Ultra has up to 128 GB unified memory and a petaflop of AI compute.
RTX Spark laptops from Lenovo, Asus, Dell, MSI, HP, and Surface are available for pre-order.

31:50Fireside chat with Nadella and Huang

What did Satya Nadella and Jensen Huang discuss on stage?

They reflected on their decades-long partnership, from Windows 95 and DirectX to CUDA and the InfiniBand supercomputer that enabled OpenAI's GPT training. They discussed MXC as a new OS primitive for agent security, hybrid intelligence as the future of computing, and RTX Spark as a complete reinvention of the personal computer that runs DirectX, OpenGL, and CUDA natively.

Nvidia was founded because of Windows; DirectX and CUDA grew from that partnership.
MXC is called a fundamental new operating system primitive for agents.
Huang says RTX Spark took over 4,000 engineering years and reinvented the PC.

Notes

Hybrid Intelligence on Windows

Microsoft's vision combines models, runtime, and silicon so AI runs locally, stays responsive, and leaves room for everything else on the PC.
GitHub Hydro Fusion routes each task to the right model, now extended to tap local models on Windows.
MAI code 1.1 Flash, a 130 billion parameter coding model, was quantized to 3 bits, reducing size nearly 80% while preserving coding quality, with a 256K context window on device.
Nvidia's new Neatron model, over 70 billion parameters, quantizes to 2-bit using just over 20 GB, ideal for RTX Spark.
DeepSeek Flash V4, a 284 billion parameter model, runs locally at 1.66 bits in about 60 GB and outperforms GPT5 on encoding and reasoning benchmarks.
Windows ML is a scalable runtime across GPUs, NPUs, and CPUs; Llama CPP is coming to Windows ML for open source models on day one.

Copilot and Developer Workflows

GitHub Copilot's auto mode picks local models for simple tasks; Kayla showed issue triage running fully offline in airplane mode.
Local sub sessions generated 1.66 million tokens for free; automations can run overnight on local models without spending credits.
Hybrid intelligence comes to the GitHub Copilot app, CLI, and VS Code starting October 15th.
The new Copilot gains local context, local actions, and local models; Autopilot demoed organizing and zipping tax files locally.

Hardware Announcements

Surface Laptop Ultra, built around Nvidia RTX Spark, offers up to 128 GB unified memory and up to a petaflop of AI compute, over 2.5x the thermal capacity of existing laptops, and over 99% of power on battery.
It features magnetic connect, the world's first built-in magnetic USB-C charging, and the brightest peak HDR display of any laptop.
RTX Spark laptops from Lenovo, Asus, Dell, MSI, HP, and Surface are available for pre-order, shipping next Friday; Call of Duty comes to RTX Spark in 2027.
DGX Station for Windows runs frontier-class models like Llama V4 Maverick, Kimmy K2.6, and DeepSeek V4 Pro at over a trillion parameters.

Nadella and Huang Fireside

Huang said Nvidia was founded because of Windows; DirectX and programmable shaders co-invented with Microsoft evolved into CUDA.
Nadella credited Huang's consistent vision of accelerated computing; their InfiniBand HPC work in Azure enabled the supercomputer used to train GPT for OpenAI.
MXC is described as a fundamental new OS primitive for agents, providing containment, identity, observability, and governance.
Huang said RTX Spark took over 4,000 engineering years and reinvented the PC, running DirectX, OpenGL, and CUDA natively for the first time.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.