Run an LLM locally on your Windows PC

Why run an LLM on your PC

The model answers and reads your files on the PC. It works without internet. There is no subscription, no quota, and no account needed. The trade-off: a model that fits a PC is smaller than the largest online models — excellent for writing, summaries and your documents, less so for the hardest reasoning.

What you need

Windows 10 or 11, 64-bit. Two memories count. On the processor, the computer's memory (RAM) decides: 8 GB for about 4 billion parameters, 16 GB for 8 billion, 24 GB for 14 billion. On a graphics card, its own memory (VRAM) decides: 8 GB for models of 4 to 8 billion parameters, 16 GB for 14 billion or gpt-oss-20b, 24 to 32 GB for 32 billion. A model that fits the graphics card writes much faster than on the processor.

Which model to choose

Qwen3.5 writes well in many languages at every size; Llama, Gemma, Mistral and gpt-oss are good alternatives. With a graphics card, choose a model that fits its memory entirely: one that spills into the computer's memory slows down a lot. Without one, stay with the small models, of about 4 billion parameters.

Three ways to do it

• Twoody — the same app as on the Mac: it suggests the model for your PC and installs it in one click, reads your folders of documents and cites them. Its installer is signed by Osmove; the model runs on the graphics card through Vulkan when the driver allows, otherwise on the processor.
• LM Studio — an app to explore thousands of models and their settings. An x64 PC whose processor supports AVX2, or an ARM PC (Snapdragon X Elite); 16 GB of memory recommended.
• Ollama — an engine used from the terminal, an API or its app. Windows 10 22H2 or later; it uses NVIDIA cards, and AMD Radeon cards through ROCm or Vulkan.

What differs from the Mac

It is the same app, with the same conversations and documents. Three things are Mac only: reading scanned PDFs, Contacts and Calendar, and dictation by macOS speech recognition — on Windows, your voice is transcribed on the PC by whisper.cpp, which ships with the app. As on the Mac, Twoody updates itself, after your click.

With Twoody, in three steps

1. 1

   Join the waitlist: Twoody is in private beta, on Windows as on the Mac and Linux, and we will write to you when it opens to you.

2. 2

   Install the model it suggests for your PC: 2.8 or 5.8 GB, one click.

3. 3

   Ask your first question — and turn Wi-Fi off if you like.