Microsoft shows hybrid AI running large models locally on Windows

Microsoft has demonstrated hybrid intelligence on Windows that runs large AI models locally while switching to cloud models when needed.

Reportedly, three models were shown running directly on Windows PCs. MAI Code 1.1 Flash reportedly has 130 billion parameters and was quantized to 3 bits, cutting its size by nearly 80 percent while keeping code quality. A new Nvidia Nemotron model reportedly exceeds 70 billion parameters, uses 2-bit quantization, and occupies a little more than 20 GB of memory. DeepSeek V4 Flash, reportedly the largest at 284 billion parameters, was compressed to 1.6 bits and fits in 60 GB of memory.

Hybrid intelligence reportedly lets the system choose between local and cloud models. The local model handles tasks it can manage, and the system routes more demanding work to cloud services such as GPT, Claude, or Gemini. Microsoft said customers want to stretch token budgets without losing access to the strongest models.

Local execution keeps data on the device and removes token costs. Work can continue offline, and queries no longer draw from paid token quotas. DeepSeek V4 Flash is reportedly a mixture-of-experts model with only 13 billion parameters active per token, which limits compute needs even at its full size.

Target hardware centers on Nvidia RTX Spark chips. These chips support up to 128 GB of unified memory. Microsoft highlighted the Surface RTX Spark Dev Box as an example machine built for local AI workloads.

Microsoft presented the models at its October 7 Windows and Surface event. The session focused on AI agents and showed the local models in action.