Microsoft is enabling advanced DeepSeek and Nvidia models to run locally on Windows computers to enhance independence and fast performance.
- Microsoft’s announcement and the shift toward running open models locally
- DeepSeek V4 Flash specifications and ultra-compression technology
- Nvidia NeMoTron models and Windows ML ecosystem integration
- Modern Surface computers and the challenges of compressed model quality
- Frequently asked questions
Microsoft’s announcement and the shift toward running open models locally
Microsoft announced on Wednesday, October 7, its intention to make open-weight artificial intelligence models developed by DeepSeek and Nvidia available to run directly and locally on Windows operating system computers. This prominent announcement came during the company’s Windows and Surface device conference held in San Francisco, its first in-person event dedicated to laptops in over two years.
This step forms a core pillar of the company’s strategy called “hybrid intelligence,” which is based on distributing computing workloads so that models run locally on the devices’ internal processors whenever possible, while falling back to the cloud to process highly complex operations. Pavan Davuluri, head of Windows and Surface devices, emphasized in statements reported by CNET that local artificial intelligence capabilities “will reshape the next chapter in the history of personal computers and open unprecedented horizons for productivity.”
DeepSeek V4 Flash specifications and ultra-compression technology
Official presentations on the conference stage revealed a technical surprise: the ability of the “DeepSeek V4 Flash” model, a massive system comprising 284 billion parameters, to operate efficiently within random access memory of no more than 60 gigabytes, after being subjected to highly complex software compression and quantization processes that reduced data size to just 1.6 bits per calculation parameter.
Microsoft officials described this achievement as delivering “intelligence that rivals leading models within 60 gigabytes of memory.” Meanwhile, Nvidia explained that this system relies on a mixture-of-experts architecture, activating only 13 billion active parameters while processing each query. Reuters reported that this model is capable of outperforming the “GPT-5” model in certain specialized tasks related to coding and mathematical reasoning.
Nvidia NeMoTron models and Windows ML ecosystem integration
The announcement also included the revelation that Nvidia plans to launch a new local model from the NeMoTron family on October 15, featuring more than 70 billion parameters and boasting advanced capabilities in running autonomous agents. SlashGear confirmed that both models, DeepSeek and NeMoTron, will operate in full compatibility with the Windows ML software execution environment.
Microsoft also revealed the integration of the widely popular open-source inference engine “llama.cpp” into Windows ML libraries, equipping it with advanced application programming interfaces (APIs) to simplify the task of developers in building generative artificial intelligence applications without complexity. Accompanying announcements included providing a local version of the “MAI Code 1.1 Flash” model, updating the “Hydra Fusion” tool on the GitHub platform to automatically distribute programming tasks between the cloud and local processors, and standardizing the “MXC” secure execution containers.
Modern Surface computers and the challenges of compressed model quality
These advanced models target the high-end personal computer category equipped with Nvidia’s “RTX Spark” chip, which supports unified memory of up to 128 gigabytes. The new “Surface Laptop Ultra,” starting at $2,599, serves as a flagship model for this category, upon which companies are betting to move programming work and content generation outside of cloud data centers.
High compression technology at a rate of 1.6 bits is considered the actual engineering key to achieving this leap; without it, the model at the 4-bit level would have required about 142 gigabytes of memory, exceeding the maximum capacity of available devices. However, technical experts cautioned that extreme weight compression can sometimes lead to an unexpected decline in output quality, emphasizing that claims of “leading intelligence” represent the company’s internal evaluation and have not yet undergone independent testing and auditing by neutral evaluation bodies.
Frequently asked questions
Question: What are the open-source models Microsoft announced it is bringing to Windows?
Answer: They include the massive DeepSeek V4 Flash model and the high-capability NeMoTron model developed by Nvidia.
Question: How can a model with 284 billion parameters run in 60 gigabytes of memory?
Answer: Thanks to software compression and quantization technology down to 1.6 bits per parameter, which reduced its operational size to fit the computers’ random access memory.
Question: What is the role of the llama.cpp engine in the Windows ML ecosystem?
Answer: It was integrated to provide standardized, fast APIs allowing developers to build local artificial intelligence applications directly on Windows.
Question: What are the hardware specifications capable of running these models locally?
Answer: They require computers equipped with powerful GPU chips, such as the Nvidia Spark chip, and unified memory starting at 60 gigabytes and reaching up to 128 gigabytes.
