NPU vs GPU: which chip runs local models in 2026

If your new laptop has an NPU, it is natural to expect it to run local models. The short answer on October 14, 2026: an NPU can run language models, but only through software written for it, and the most popular local models tools still use the GPU. This guide explains what each chip is for, which programs actually use the NPU, the limits the vendors themselves document, and how to decide what to buy. We read Microsoft's, AMD's, Intel's, Ollama's, llama.cpp's and LM Studio's documentation. We have not benchmarked an NPU: our test machine (Intel Core i9-12900KF with an RTX 4070) does not have one, so there are no speed figures here that we did not measure.

NPU vs GPU in one table

Table: NPU and GPU for local models (vendor documentation, checked October 14, 2026)

What an NPU is and what TOPS means

A neural processing unit is a chip dedicated to the arithmetic of neural networks. Microsoft calls Copilot+ PCs "a new class of Windows 11 hardware" with an NPU that "can perform more than 40 trillion operations per second (TOPS)", and many Copilot+ features in Windows require that 40 TOPS minimum. The first devices used Qualcomm's Snapdragon X Elite, and Microsoft also lists AMD Ryzen 300-series (Strix Point) and Intel Core Ultra 200V processors.

Microsoft's support page claims the NPU uses less power and is far more efficient than a CPU or GPU at running models, and frees the CPU and GPU for other work. Treat that as the vendor's description of the design goal, not as a measurement. Two technical points in Microsoft's developer guidance matter more for local models:

1. The NPU needs software written for it. "Like other hardware resources on a PC, the NPU needs software to be specifically programmed to take advantage of the benefits it offers." An app that was not built for the NPU will not use it.
2. Models usually have to be converted. "Many NPU devices, however, only support integer math in lower bit format, such as INT8", so models "need to be converted (or quantized) to run on the NPU". Our guide to what quantization is explains what that costs in quality.

TOPS also says nothing about how much memory the model can use. For language models, memory is usually the first limit you hit, which is why our guide to how much VRAM an LLM needs starts there.

Which programs use the NPU today

This is where the gap is clearest. According to each project's own documentation:

• Ollama lists NVIDIA GPUs (compute capability 5.0 and newer), AMD GPUs through ROCm, Apple GPUs through Metal, and other GPUs through Vulkan. There is no NPU in its hardware support page. If you are trying to get Ollama onto your graphics card, see Ollama not using your GPU.
• LM Studio runs on Windows on x64 and ARM (Snapdragon X Elite) and recommends at least 16 GB of RAM and 4 GB of dedicated VRAM. Its system requirements do not mention the NPU.
• llama.cpp, the engine under Ollama (our Ollama vs llama.cpp comparison explains the relationship), lists NPU backends separately from the usual ones: Hexagon for Snapdragon, CANN for Huawei's Ascend NPU, and OpenVINO for Intel CPUs, GPUs and NPUs, which its README marks as "in progress". The Snapdragon backend is built from source with Qualcomm's toolchain, so it is not what you get from a normal download.

The NPU path runs through each chip maker's own stack instead:

• AMD's NPU software for Ryzen runs language models "NPU-only" or in a "hybrid" mode that splits work between the NPU and the integrated GPU, through ONNX Runtime GenAI, on Ryzen 300-series (Strix Point) processors. On Ryzen 7040 and 8040 (Phoenix and Hawk Point), AMD lists only GPU and CPU for language models. AMD also points to FastFlowLM, an NPU-first runtime, and to Lemonade Server, which offers an API that apps written for Ollama or OpenAI can call, with a set of pre-optimized models such as Llama, Mistral, Qwen, Gemma and Phi.
• Intel runs language models on its NPU through OpenVINO GenAI. Models must be exported with 4-bit weights (INT4 or NF4).
• Microsoft recommends Windows ML, which picks an execution provider for the hardware: QNN for Qualcomm NPUs, OpenVINO for Intel and VitisAI for AMD, falling back to the GPU or CPU. These providers download on demand on Windows 11 24H2 and later. Foundry Local, Microsoft's on-device runtime for apps, selects the GPU or NPU automatically and offers a curated catalog that includes GPT OSS, Qwen, DeepSeek, Mistral, Phi and Whisper.

The limits vendors document for NPUs

The NPU route works, but the documentation shows trade-offs you will not see on a spec sheet:

• Short default context on Intel. OpenVINO's NPU pipeline "supports input prompts up to 1024 tokens" by default and guarantees at least 128 tokens of response, because it uses fixed shapes. You can raise the limit, and dynamic prompts arrived in OpenVINO 2025.3, but the defaults tell you what the hardware was tuned for. Our guide to context windows explains why 1,024 tokens is short for a chat.
• Memory still matters. Intel warns that on Core Ultra Series 2 systems "more than 16GB of RAM may be required" for prompts over 1,024 tokens with models above 7 billion parameters.
• Curated model lists. AMD, Intel and Microsoft all ship lists of models already converted for their NPUs. Foundry Local says its catalog is "intentionally curated" and not meant for "general-purpose model experimentation".
• Driver versions. Windows ML lists minimum NPU driver versions for Qualcomm and a supported driver range for AMD's provider, so an out-of-date driver can silently send the work back to the CPU.

Which should you buy for local models?

• You want to chat with open models in Ollama or LM Studio: prioritize a GPU with enough VRAM. Our best GPU for local LLMs guide covers the options.
• You want long battery life and the Copilot+ features in Windows: a Copilot+ PC with a 40+ TOPS NPU is what those features require.
• You are a developer shipping a model inside an app for many laptops: the NPU is worth targeting through Windows ML or Foundry Local, because the runtime picks the best chip on each user's machine.
• You already own a Ryzen 300-series (Strix Point) or Core Ultra laptop: try the vendor stack (Lemonade or FastFlowLM on AMD, OpenVINO on Intel) with one of its pre-converted models, and compare it with the same model on the GPU in LM Studio on your own prompts.

If you are new to all of this, start with running an LLM locally on whatever hardware you have, and only then decide whether the NPU is worth the extra setup.