Microsoft and NVIDIA bring local agents to Windows with MAI-Code-1.1 Flash and RTX Spark

Microsoft called it 'a new chapter for Windows.' Nadella wrote on LinkedIn (11,868 likes) about bringing 'unmetered intelligence to every desk and every home' and making every PC a place where agents can work securely. The centerpiece is MAI-Code-1.1 Flash, a 137B-parameter coding model with a 256K-token context window, optimized for local execution. NVIDIA's RTX Spark platform provides the hardware anchor. Huang remarked that NVIDIA was founded because of Windows, and that agents are now coming to it.

The new 'Hybrid Intelligence' approach splits work between cloud and device. GitHub Copilot can offload tasks to the local MAI model to cut cloud costs 'without sacrificing quality,' Microsoft says. Copilot can also take action directly on the PC, keeping sensitive work on the device. That puts model routing and cost control into the OS layer.

Competitively, this is Microsoft hedging its frontier-model dependence. MAI is an in-house family, while Foundry hosts GPT-6 and Claude Opus 5.5. Local inference also feeds the cost narrative: Microsoft is shifting Copilot to pay-per-use. Apple, meanwhile, is tightening macOS permissions for agents, a very different philosophy on local agent access.

What to watch: real hardware requirements (a 137B model on a PC implies substantial memory, likely high-end RTX Spark systems), independent coding benchmarks against cloud models, and security. A 28-second prompt-injection chain affecting Copilot CLI circulated this week, a reminder that agents acting on the desktop widen the attack surface.