The verdict in 30 seconds
LM Studio can run a downloaded model locally. Working locally does not mean that every feature of the app is offline: consult the software settings and privacy notice.
✅ When it works well
- CPU and system supported by the chosen version
- Enough RAM for the model, context and applications
- Optional GPU, if supported by drivers and backend
- Disk space for the app and model files
❌ When it doesn't start
- Windows x64 without AVX2: CPU requirement not met; on Linux check the build and runtime
- Not enough memory for the chosen model and context
- Unsupported operating system version
- Drivers or backend incompatible with the GPU
The good news
A dedicated GPU is not required if the CPU and runtime are compatible. Performance and memory depend on model, quantization, context and hardware; 8 GB of RAM can be tight and it's worth trying a small model first.
CPU requirement: AVX2 on x64
The CPU is one of the requirements to check along with the operating system, memory and graphics backend.
For Windows x64, the official LM Studio requirements indicate AVX2. On Linux x64 the package includes AVX2 by default. A CPU with only AVX does not meet the Windows requirement; for Linux check the package, runtime and available instructions. Apple Silicon uses the native ARM path and requires macOS 14 or later. Check the updated official requirements before installing.
| CPU family | Indication | AVX2 | Result with the current x64 version |
|---|---|---|---|
| Windows x64 without AVX2 | different models | no | Does not meet the indicated CPU requirement |
| Linux x64 without AVX2 | check the build | to be checked | The package includes AVX2 by default; check whether the chosen build is compatible |
| x64 CPU with AVX2 | check the model | yes | Meets the CPU requirement; also check the other requirements |
| Apple Silicon | M series | ARM | Uses the native path for Apple Silicon |
If your x64 processor doesn't have AVX2
If the CPU does not meet the requirement, consider these alternatives:
1. Check whether the runtime offers a compatible backend: availability depends on the version and the platform.
2. Use a remote model: the PC can act as a client to a service or another computer.
3. Try a very small model first only if you found a maintained build with compatible requirements. Performance varies and is not guaranteed.
Configurator: the settings for your PC
Choose your machine: the configurator estimates a suitable model, an initial value for the context and the possible use of the GPU. Always check free memory and the runtime requirements.
Step-by-step installation
- Download the correct installer
Go to lmstudio.ai/download. Choose the package for your supported platform and check the system requirements published by LM Studio.
- Before opening everything, free up RAM
Close applications you don't need and check how much memory is free. System usage changes based on version, drivers and running programs.
- Open the search and filter by RAM
In the LM Studio search field, each model shows its file size and required RAM. Adjust the sidebar to the GB actually free, not the installed total.
- Download just one model, in Q4_K_M
Choose a quantization based on memory and quality. Q4_K_M is a common format, but the impact varies by model and task. With little RAM try a small variant first and check the required context.
- Configure before writing
In the model panel, start with a modest context. If the GPU and backend are supported, try a partial offload leaving room for cache and buffers. Keep the default thread value, then compare results.
- Test a real request
Use a task you care about and check quality, speed and memory usage. Performance and answers vary with model, quantization and configuration.
Where the models end up
Windows: C:\Users\[tuouser]\.cache\lm-studio\models
Linux: ~/.lmstudio/models
macOS: ~/Library/Application Support/LM Studio/models
The model path and management can vary with installation and version; use the app's settings section to find it and remove files.
GPU: NVIDIA, AMD and integrated
A GPU can speed up inference if supported by drivers, runtime and backend. VRAM sets a limit on the part of the model that can be loaded onto the card.
| Card | VRAM | Real effect | Recommended? |
|---|---|---|---|
| Integrated graphics only | shared | Acceleration varies based on GPU, drivers and backend | Depends |
| GTX 1050 Ti / GTX 1650 | 4 GB | The model can be split between VRAM and RAM if the runtime supports offloading | depends |
| GTX 1660 Super, RTX 2060 | 6 GB | VRAM limits the offload; part of the model can stay in RAM | so-so |
| Dedicated GPU with 8 GB | 8 GB | The capacity can suffice for some small/medium quantized models; context and runtime use additional memory | depends |
| Dedicated GPU with 12 GB | 12 GB | More room for weights and context; check the size and backend of the chosen model | depends |
| Dedicated GPU with 24 GB | 24 GB | Can host larger models, but the required capacity depends on quantization and context | depends |
| AMD or Intel GPU | variable | Compatibility and performance depend on model, system, drivers and supported backend | check |
The parameter that matters is GPU Offload
In LM Studio, the GPU Offload field indicates how many layers of the model are assigned to the GPU. Leave headroom in VRAM for context, cache and buffers; partial offloading can be useful if the model doesn't fit entirely. The available options and the result depend on the backend.
Common problems and how to fix them
«Failed to load model» or the app closes immediately
It can depend on insufficient memory, the runtime or required CPU instructions, an incomplete file, or the GPU backend. Check the error message, the free RAM and the requirements of the installed version. If it reports AVX2, compare the CPU model with the current requirements.
The model loads but answers are extremely slow
Check the context and reduce it if memory is at its limit; try different thread values because the best one varies with CPU and runtime; observe whether the model is loaded in RAM, VRAM or both; check quantization and backend. Loading times depend on the disk and the file size.
Answers are good but slow to grow (character by character)
The text is generated one token at a time. The pace depends on the CPU or GPU, model size, quantization, context and runtime; a compatible GPU can help, but it isn't the only factor.
«Context length exceeded» or the conversation gets stuck
You're asking the model to remember more than you gave it. The practical context depends on the model and free memory; try reducing it if the error appears. If you need to remember whole documents, load them as files and ask questions about them, instead of pasting everything into chat.
On AMD the GPU is not used
Check that the GPU, driver version and backend are supported by the version of LM Studio you use. Look in the log to see whether the runtime actually selected the GPU and try a configuration known to be compatible.
Everything is configured correctly but it's still slow
Check whether the memory runs in single or dual channel. Dual channel increases theoretical bandwidth and can help CPU inference, but the gain varies with processor, memory, model and runtime and doesn't necessarily equal double.
Frequently asked questions
Does LM Studio work on a PC from 2010?
For Windows x64, LM Studio lists AVX2 in the requirements. The Linux x64 package includes AVX2 by default: check the build and the CPU's instructions. Apple Silicon uses a native ARM version. Check the official requirements before installing.
Why does it say instructions are missing?
Check the CPU requirements, error message, operating system, runtime and app version. On Windows x64, AVX2 is listed among the requirements; on Linux x64, check the instructions required by the current build. An alternative build must be verified.
How large should the models be for 8 GB of RAM?
With 8 GB total it's prudent to try smaller models and reduced contexts. A quantized 7-8B can be at the limit after considering the system and cache; a 14B often requires more free memory. Check the model and the memory actually available.
Is it worth using the integrated graphics?
It depends on the integrated graphics model, the drivers and the supported backend. It usually uses RAM shared with the system; a dedicated GPU can offer more bandwidth and VRAM, but the result is not guaranteed by the memory figure alone.
How do I check whether my processor has AVX?
On Windows use CPU-Z and look at the Instruction Set line. On Linux check the flags in /proc/cpuinfo. For Windows x64, AVX2 is listed in the requirements; on Linux x64, compare the exposed instructions with those required by the current build.
Is LM Studio or Ollama better?
Interface and features are different; performance and compatibility depend on version, model, backend and hardware. LM Studio offers a graphical interface; Ollama exposes an HTTP API. The full comparison is on the Ollama vs LM Studio page.