LM Studio is a free desktop app for downloading and chatting with large language models on your own PC. Our Ollama and LM Studio install guide covers getting the app installed. This page is about the next step: picking a model file that will actually fit, and loading it without starving the rest of Windows of memory.
Forge PC has not benchmarked models or measured tokens per second in LM Studio. The steps and limits below come from LM Studio's official documentation, read on 13 November 2026. Menu names can change between app versions.
What you need first
LM Studio's Windows requirements are:
An x64 CPU with AVX2, or a Windows-on-Arm PC such as a Snapdragon X Elite system.
At least 16GB of system RAM, recommended. The app does not call this a hard minimum, but it is honest about how much memory models use.
At least 4GB of dedicated VRAM, recommended.
On a Mac, LM Studio needs Apple Silicon and macOS 14 or newer; Intel Macs are not supported. On Linux it ships as an AppImage for x64 and Arm64, with Ubuntu 20.04 or newer required, and versions newer than Ubuntu 22 "not well tested", in the developer's words.
If you are still sizing a PC for this, our guides to how much RAM local AI needs, how much VRAM you need and the best GPU for local models are the capacity side of the same problem. This page assumes the app is already installed.
Step 1: Find a model in Discover
LM Studio has a built-in downloader that fetches supported models from Hugging Face. Open the Discover tab. Official documentation says you can jump to it from anywhere with Ctrl + 2 on Windows or Linux, or Cmd + 2 on a Mac.
You can search by a keyword such as llama, gemma, qwen or gpt-oss, by a user/model string, or by pasting a full Hugging Face URL. Stick to model pages you recognise, and prefer files the app itself marks as compatible.
Step 2: Choose a quantisation you can actually run
For any given model you will usually see several download options with names such as Q3_K_S, Q4_K_M or Q8. LM Studio's docs say these are copies of the same model at different levels of quantisation: a compression method that shrinks the file and, in return, gives up some fidelity.
The same documentation's advice is to choose a 4-bit option or higher if your machine can run it. That is a useful starting rule. A higher-bit file (Q6, Q8) is closer to the original weights and is also larger, so it needs more RAM and VRAM. A lower-bit file may load on a smaller PC and then disappoint you, or may not be offered as a good default.
If the letters and numbers still look opaque, our explainer on LLM quantisation and GGUF walks through the format these files usually use. LM Studio can also run MLX models on Apple Silicon; on a typical Windows gaming PC you will be looking at GGUF.
Do not start with the largest model on the list. A 7- or 8-billion-parameter model at Q4 is a safer first download than a 70-billion-parameter file that will spill into system RAM and crawl. Disk space matters too: these files are several gigabytes each, and you can change LM Studio's models folder from the My Models page if your system drive is tight.
Step 3: Load it, then set context and GPU offload
After the download finishes, open the model loader (the control that puts a model into memory for a chat or for the local server). This is where two settings matter:
Context length. This is how much of the conversation and any attached text the model can consider at once. A longer context uses more memory. Start with the app's default, and only raise it if you are hitting the limit in a long chat or when reading a large document.
GPU offload. This is how many layers of the model go on the graphics card. More offload is usually faster, until you run out of VRAM, at which point the PC will spill into system RAM or fail to load. If a model will not load, reduce offload or pick a smaller quantisation.
LM Studio's documentation also describes per-model defaults, so you can save a context size and offload setting for a model you use often, and a command-line tool, lms load, with flags such as --context-length and --gpu (from off, through a 0–1 fraction, to max). If you are only using the app, the on-screen loader is enough.
The docs say LM Studio will warn you before loading a model that is likely to run out of memory. Take that warning seriously. Closing a failed load and choosing a smaller file is faster than letting Windows start paging.
If you use the lms command line that ships with the app, lms load --estimate-only <model> prints a memory estimate without loading the model. Official flags also include --context-length and --gpu (off, a 0–1 fraction, or max).
Step 4: Chat, or turn on the local server
Once a model is loaded, you can chat with it in the app. LM Studio can also run as a local server, typically at http://localhost:1234, with OpenAI-compatible endpoints, so other software on the same PC can use the loaded model. If you turn on access from other machines on your network, LM Studio's documentation recommends enabling authentication.
You can use the app entirely offline once the model files are on disk. That is one of the reasons to run models locally in the first place, alongside the points in our cloud AI vs local LLM comparison.
Our verdict
In LM Studio, open Discover (Ctrl + 2), pick a model you have heard of, and download a 4-bit or higher GGUF if your PC has the memory. Load it with a modest context length and as much GPU offload as your VRAM will take, and drop to a smaller file if the app warns that you will run out of memory. A small model that actually fits will serve you better than a huge one that swaps to disk. We have not timed these models ourselves; the steps come from LM Studio's official documentation.
