How much RAM or VRAM does LM Studio need?

LM Studio itself is light; the memory goes to the model you load. At 4-bit a 3B model needs under 3GB, an 8B around 6GB and a 70B over 40GB. The table on this page gives the figure for every model, including the KV cache and framework overhead rather than the weights alone. LM Studio shows a compatibility indicator against your own machine when you browse models, which is the feature most people install it for.

Can LM Studio run models larger than my VRAM?

Yes, by offloading some layers to system RAM, and the GPU offload slider in the app controls how many. The layers that stay on the CPU run at system memory bandwidth instead of GPU bandwidth, so throughput drops sharply. It is the right tool for occasionally reaching one tier above your hardware, not for daily use.

It is free to download and use. It is closed source, unlike Ollama and llama.cpp, so if you need an open-source tool or have questions about commercial use inside your organisation, check its current terms rather than assuming.

Should I use GGUF or MLX in LM Studio?

Only Apple Silicon users have this choice. GGUF works everywhere and exists for every model. MLX is Apple's own framework, built around unified memory, and community reports consistently favour it for generation speed on a Mac — though that varies by model and is not something we have measured. Availability usually decides it: GGUF covers the long tail, MLX covers the popular models.

They run the same models on the same hardware, so it is not a capability question. LM Studio is a desktop application with a model browser, a chat window and visible controls, which suits you being the one talking to the model. Ollama is a background service with a command line and an HTTP API, which suits another application calling it. Installing both is common.

LM Studio models and VRAM requirements

An application. Browse, download, chat, inspect — no terminal needed. Below is every model in our directory with the memory it needs at Q4 and Q8, and the lms command that fetches it — so you can size a download before you start it. Of the 75 machines we profile, 68 can run it.

All 30 models, smallest first

LM Studio has a built-in Hugging Face browser, so the simplest route is to search any of these by name in the app — the CLI column is the equivalent for people who prefer a terminal (lms get downloads by name, lms load addresses a repo directly). Q4 is the usual starting point; Q8 roughly doubles the memory for a quality difference most people cannot pick out in ordinary use.

Model	Q4 VRAM	Q8 VRAM	CLI command
Meta Llama 3.2 3B Instruct 3.21B · General	2.5GB	4.2GB	`lms load meta-llama/Llama-3.2-3B-Instruct-GGUF`
Qwen 2.5 Coder 7B Instruct 7.6B · Coding	5.5GB	9.5GB	`lms load Qwen/Qwen2.5-Coder-7B-Instruct-GGUF`
DeepSeek R1 Distill Llama 8B 8.0B · Reasoning	5.8GB	9.8GB	`lms load deepseek-ai/DeepSeek-R1-Distill-Llama-8B-GGUF`
Llama 3.1 8B Instruct 8.03B · General	6GB	10GB	`lms load meta-llama/Meta-Llama-3.1-8B-Instruct-GGUF`
Qwen3 8B 8.2B · General	6.4GB	10.7GB	`lms get qwen3-8b`
Google Gemma 2 9B 9.24B · General	7GB	11.5GB	`lms load google/gemma-2-9b-it-GGUF`
Gemma 4 12B 12B · General	8.7GB	14.7GB	`lms get gemma-4-12b`
Mistral NeMo 12B Instruct 12.2B · General	9GB	15GB	`lms load mistralai/Mistral-Nemo-Instruct-2407-GGUF`
DeepSeek R1 Distill Qwen 14B 14.8B · Reasoning	10.5GB	17.5GB	`lms load deepseek-ai/DeepSeek-R1-Distill-Qwen-14B-GGUF`
Qwen 2.5 Coder 14B Instruct 14.7B · Coding	10.5GB	17GB	`lms load Qwen/Qwen2.5-Coder-14B-Instruct-GGUF`
Microsoft Phi-4 14B 14.7B · Reasoning	10.5GB	17.5GB	`lms load microsoft/phi-4-GGUF`
Mistral Codestral 22B 22.2B · Coding	15GB	26GB	`lms load mistralai/Codestral-22B-v0.1-GGUF`
OpenAI gpt-oss-20b 20.9B (3.6B active) · Reasoning	15.3GB	24.2GB	`lms get openai/gpt-oss-20b`
Mistral Small 3 24B 24.0B · General	16GB	28GB	`lms load mistralai/Mistral-Small-24B-Instruct-2501-GGUF`
Gemma 4 26B-A4B (MoE) 26.5B (3.8B active) · General	17.4GB	30.1GB	`lms get gemma-4-26b-a4b`
Google Gemma 2 27B 27.2B · General	18GB	32GB	`lms load google/gemma-2-27b-it-GGUF`
Qwen3.8 27B 27.8B · General	18.2GB	31.5GB	`lms get qwen3.8-27b`
Qwen3-Coder 30B-A3B Instruct 30.5B (3.3B active) · Coding	19.8GB	34.3GB	`lms get qwen3-coder-30b`
Gemma 4 31B 31.3B · General	20.3GB	35.2GB	`lms get gemma-4-31b`
DeepSeek R1 Distill Qwen 32B 32.5B · Reasoning	21.5GB	38GB	`lms load deepseek-ai/DeepSeek-R1-Distill-Qwen-32B-GGUF`
Qwen 2.5 Coder 32B Instruct 32.5B · Coding	21.5GB	38GB	`lms load Qwen/Qwen2.5-Coder-32B-Instruct-GGUF`
Cohere Command R 35B 35.0B · RAG	23GB	40GB	`lms load CohereForAI/c4ai-command-r-v01-GGUF`
Qwen3.6 35B-A3B (MoE) 36B (3B active) · General	23.1GB	40.2GB	`lms get qwen3.6-35b-a3b`
Llama 3.3 70B Instruct 70.6B · General	42GB	78GB	`lms load meta-llama/Llama-3.3-70B-Instruct-GGUF`
DeepSeek R1 Distill Llama 70B 70.6B · Reasoning	45GB	80GB	`lms load deepseek-ai/DeepSeek-R1-Distill-Llama-70B-GGUF`
Qwen 2.5 72B Instruct 72.7B · General	45GB	82GB	`lms load Qwen/Qwen2.5-72B-Instruct-GGUF`
OpenAI gpt-oss-120b 116.8B (5.1B active) · Reasoning	66.5GB	125.8GB	`lms get openai/gpt-oss-120b`
Llama 4 Scout (17B active, 16 experts) 109B (17B active) · General	66.9GB	117.5GB	`lms get llama-4-scout`
DeepSeek V4-Flash 284B (13B active) · Reasoning	171.9GB	303GB	`lms get deepseek-v4-flash`
DeepSeek V3 671B MoE 671B (37B active) · General	400GB	740GB	`lms load deepseek-ai/DeepSeek-V3-GGUF`

Figures are the minimum to load at that quantization. Long context adds to them — the VRAM calculator splits the requirement into weights, KV cache and overhead for any context length.

On a Mac: the MLX option

This is the one thing LM Studio does that Ollama does not. Alongside GGUF it can run MLX, Apple's own array framework, which was written for unified memory from the start. It matters here because 24 of the 75 machines we profile are Macs.

GGUF

Universal. Every model, every runner, every platform.

MLX

Apple-native. Often faster on a Mac, but Mac-only and less complete.

Community reports consistently favour MLX for generation speed on Apple Silicon, but that varies by model and we have not measured it — we have no hardware lab. Availability is usually the practical decider. GGUF vs MLX in full, or the Mac VRAM optimizer for how much unified memory your GPU can address.

What fits your machine

Or start from a specific machine in the hardware matrix, or from a card in GPU comparison.

LM Studio or Ollama?

Both run the same GGUF models on the same hardware, so this is not a capability comparison. Ollama is a background service with a command line and an HTTP API, and it is what you want if something else in your stack is going to call the model. LM Studio is a desktop application with a model browser, a chat interface and a server mode you can switch on, and it is what you want if you are the one talking to the model. Many people install both.

Ollama vs LM Studio in full Ollama models and VRAM New to local models? Start here

Common questions

How much RAM or VRAM does LM Studio need?: LM Studio itself is light; the memory goes to the model you load. At 4-bit a 3B model needs under 3GB, an 8B around 6GB and a 70B over 40GB. The table on this page gives the figure for every model, including the KV cache and framework overhead rather than the weights alone. LM Studio shows a compatibility indicator against your own machine when you browse models, which is the feature most people install it for.
Can LM Studio run models larger than my VRAM?: Yes, by offloading some layers to system RAM, and the GPU offload slider in the app controls how many. The layers that stay on the CPU run at system memory bandwidth instead of GPU bandwidth, so throughput drops sharply. It is the right tool for occasionally reaching one tier above your hardware, not for daily use.
Is LM Studio free?: It is free to download and use. It is closed source, unlike Ollama and llama.cpp, so if you need an open-source tool or have questions about commercial use inside your organisation, check its current terms rather than assuming.
Should I use GGUF or MLX in LM Studio?: Only Apple Silicon users have this choice. GGUF works everywhere and exists for every model. MLX is Apple's own framework, built around unified memory, and community reports consistently favour it for generation speed on a Mac — though that varies by model and is not something we have measured. Availability usually decides it: GGUF covers the long tail, MLX covers the popular models.
LM Studio or Ollama?: They run the same models on the same hardware, so it is not a capability question. LM Studio is a desktop application with a model browser, a chat window and visible controls, which suits you being the one talking to the model. Ollama is a background service with a command line and an HTTP API, which suits another application calling it. Installing both is common.

Guides that go deeper

Choosing a runner, picking a Mac for local models, and what each quantization costs.