LM Studio is one of the most hardware-aware LLM frontends available. Unlike tools that run the same inference backend regardless of platform, LM Studio selects its backend based on what hardware it detects: MLX on Apple Silicon, CUDA on NVIDIA, and Metal as an Intel Mac fallback. This means a Mac M5 Pro running LM Studio gets meaningfully better performance than the same hardware running a tool defaulting to llama.cpp’s CPU path.
That backend selection decision is what this guide is built around.
Quick answer: For NVIDIA desktop builds, the RTX 4090 (24GB) handles 34B models smoothly and the RTX 4060 Ti 16GB is the budget entry point for 13B at full quality. For Apple Silicon, the M5 Pro 24GB is the minimum for comfortable 13B use, and M5 Max 48GB+ handles 34B. The used RTX 3090 (24GB) remains the strongest VRAM-per-dollar option if you find one at a good price.
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB GDDR6X with full CUDA — handles 34B models in LM Studio at full precision with fast token generation. No VRAM constraints on any mainstream LLM in 2026.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
How LM Studio picks its backend
This matters because it directly affects performance, and it’s what separates LM Studio from other local inference tools.
Apple Silicon: LM Studio defaults to MLX, Apple’s native machine learning framework for Apple chips. MLX uses the unified memory architecture of M-series chips efficiently — the same memory pool serves both CPU and GPU, meaning a MacBook Pro M5 Max with 48GB has 48GB available to the model with no VRAM ceiling separate from system RAM. MLX performance on Apple Silicon is significantly faster than running llama.cpp CPU inference, and in many cases faster than GPU-offloaded llama.cpp as well.
Before LM Studio made MLX the default on Apple Silicon, tools like earlier versions of Ollama defaulted to llama.cpp — which would use CPU inference unless explicitly configured for GPU offloading. LM Studio’s automatic MLX backend is why Mac LLM performance for many users changed overnight when they switched frontends, not hardware.
NVIDIA GPUs: LM Studio uses CUDA-accelerated llama.cpp or its own CUDA inference path. Full GPU acceleration with VRAM management, quantization selection, and model splitting if needed.
Intel Mac / no supported GPU: Falls back to Metal or CPU inference via llama.cpp. Functional but significantly slower — not a recommended primary platform for LLM inference.
Before you weigh the picks below, run your own card through a quick fit check:
Before any of that matters, check the floor: LM Studio’s system requirements call for macOS 14.0 or newer on Apple Silicon, Ubuntu 20.04 or newer on Linux, and AVX2 support on x64 CPUs. That last one is worth knowing before you buy a card — a pre-2013 CPU without AVX2 will not run LM Studio no matter what GPU is in the slot.
VRAM requirements by model size in LM Studio
LM Studio’s quantization selector makes VRAM requirements variable. Here’s a practical guide to what fits where:
| Model size | Q4 quantization | Q8 quantization | Full precision (FP16) |
|---|---|---|---|
| 7B | ~4.5GB | ~8GB | ~14GB |
| 13B | ~7.5GB | ~14GB | ~26GB |
| 34B | ~20GB | ~35GB | ~68GB |
| 70B | 43GB | 75GB | 141GB |
For LM Studio on NVIDIA: if a model’s quantized size fits in VRAM, it runs fully on GPU. If it doesn’t fit, LM Studio can split layers across GPU and CPU — but layers running on CPU are dramatically slower. The practical target is fitting the entire model in VRAM for acceptable generation speed.
For Apple Silicon: unified memory means the 7B Q4 / 13B Q4 / 34B Q4 question is just about total system memory, not a separate VRAM limit. This is the architectural advantage.
For more on VRAM sizing principles, see how much VRAM do you need for local LLM.
NVIDIA picks for LM Studio
RTX 4090 (24GB) — best NVIDIA option: 24GB handles 13B models at Q8 or FP16, 34B models at Q4 and Q5, and provides fast generation on 7B models. LM Studio’s CUDA path with 24GB means no model splitting on mainstream LLMs in 2026 — everything runs fully on GPU at comfortable speeds. Community users report 25–40 tokens/second for 13B Q4 on RTX 4090, which is fast enough for productive use.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→RTX 4060 Ti 16GB — best budget 13B card: 16GB is the sweet spot for 13B model users. The RTX 4060 Ti 16GB at around $425 fits 13B Q8 (14GB) with margin, and handles 34B Q4 (20GB) with minor layer splitting. For users primarily running 7B and 13B models, this card handles LM Studio workloads well. Generation speed is slower than the 4090 due to lower bandwidth (288 GB/s vs 1,008 GB/s), but fully functional. See best GPU for 13B models for a detailed comparison.
Check NVIDIA GeForce RTX 4060 Ti 16GB on Amazon→Buy on Shopee SG→Used RTX 3090 (24GB) — best VRAM-per-dollar: If you’re willing to buy used, the RTX 3090 offers 24GB GDDR6X — the same VRAM capacity as the RTX 4090 — at significantly lower prices on the secondhand market. Generation speed is noticeably slower than the 4090 (lower memory bandwidth), but for users whose bottleneck is VRAM capacity rather than raw throughput, the 3090 gives 34B model compatibility at a fraction of 4090 pricing. LM Studio runs cleanly on RTX 3090 with full CUDA support.
Apple Silicon picks for LM Studio
The MLX backend makes Apple Silicon uniquely competitive for LLM inference in LM Studio. The math is straightforward: unified memory means no separate VRAM ceiling, and MLX performance on M-series chips is fast enough that M-series Macs can outperform lower-VRAM NVIDIA cards for certain model sizes.
M5 Pro 24GB — minimum for 13B: The M5 Pro with 24GB of unified memory handles 13B Q8 comfortably and 34B Q4 with acceptable performance. 24GB is also where the Mac mini’s M5 Pro configuration starts, which makes it the practical floor for productive 13B work; the 16GB M6 machines below it are fine for 7B but cramped for 13B Q8. LM Studio’s MLX path here gives smooth generation that would need an RTX 4060 Ti or better on the NVIDIA side.
M5 Max 48GB+ — for 34B models: 48GB of unified memory handles 34B Q8 and is the entry point for comfortable 34B use. It sits in a position no NVIDIA consumer card occupies: the RTX 4090 and 5090 stop at 24GB and 32GB, so fitting a 34B Q8 model (35GB) means a Mac or a workstation card. The M5 Max is configurable well past that — up to 128GB — but 48GB is where the capability starts and the cheapest way to reach it.
M5 Ultra — for 70B+ models: Ultra-class chips configure to 256GB and 512GB of unified memory, which runs 70B at Q8 and 34B at full precision — configurations no consumer NVIDIA GPU can reach at any price. LM Studio’s MLX backend exploits this fully. For 70B-class work locally without a multi-GPU server, this is the only consumer-accessible path. The price is workstation-level; the capability is genuine.
For a full head-to-head comparison of these platforms, see Mac vs NVIDIA for LLM.
Which GPU for LM Studio?
- You run 7B models, budget build: RTX 3060 12GB or RTX 4060 8GB handles 7B Q4/Q8 fully in VRAM. Not comfortable for 13B.
- You run 7B–13B models, NVIDIA desktop: RTX 4060 Ti 16GB (~$425) is the right call — 16GB fits 13B Q8, every 7B fits easily.
- You run 34B models, NVIDIA: RTX 4090 (24GB) or used RTX 3090 (24GB). 24GB fits 34B Q4/Q5 fully in VRAM.
- You’re on Apple Silicon, running 13B: M5 Pro 24GB minimum. 16GB is workable but cramped.
- You’re on Apple Silicon, running 34B: M5 Max 48GB+. This is the only accessible path to 34B Q8 on a single consumer device.
- You run 70B models: an M5 Ultra (256GB or 512GB) or a multi-GPU NVIDIA setup. No single consumer NVIDIA card handles 70B on its own.
- You want to explore models without committing: LM Studio’s model browser and built-in chat interface make it ideal for this. Use LM Studio for exploration, then move to Ollama for production automation.
Why LM Studio is worth using even on NVIDIA
Several GPU buyers default to Ollama because it has better automation and API support. That’s a valid workflow — but LM Studio offers something distinct that makes it worth running alongside Ollama:
Model browser: LM Studio has a built-in model discovery interface connected to HuggingFace. You can browse, filter by size and quantization, and download directly. No manual HuggingFace navigation or CLI commands.
Built-in chat interface: A polished chat UI with conversation history, system prompt editing, and context length controls. Better than Ollama’s default web UI for interactive use.
Quantization comparison: LM Studio makes it easy to test the same model at Q4, Q5, Q6, and Q8 side-by-side and assess quality vs speed trade-offs with your actual VRAM. This is valuable during the exploration phase when you’re deciding what model to run long-term.
LM Studio as exploration, Ollama for production: The common pattern among experienced local LLM users is to use LM Studio to explore new models and find quantizations that work well, then export the model path to Ollama for API-accessible, automation-friendly production use. LM Studio has an Ollama-compatible server mode that bridges this workflow. See best GPU for Ollama for Ollama-specific guidance, and best GPU for Open WebUI if you plan to put a browser chat interface in front of that Ollama backend.
LM Studio system requirements
LM Studio’s official system requirements do not name a CUDA version at all — what they do specify is AVX2 support on x64 CPUs, macOS 14.0 or newer on Apple Silicon, and Ubuntu 20.04 or newer on Linux, with 16GB of RAM and 4GB of dedicated VRAM as recommendations. On the Mac side it is still worth running a current macOS: Apple ships MLX performance work through OS updates, so the framework improves without you changing hardware.
GPU memory requirements are model-dependent — LM Studio displays available VRAM and flags whether your selected model fits before loading, which makes it more user-friendly than tools that discover VRAM limits at runtime.
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB GDDR6X on the used market at well under RTX 4090 pricing — same VRAM capacity, slower generation, but handles 34B Q4 in LM Studio without model splitting.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
For broader LLM hardware context, see how much VRAM for local LLM and best GPU for Llama 4.
Frequently Asked Questions
What are LM Studio’s GPU requirements?
LM Studio’s published requirements are macOS 14.0 or newer on Apple Silicon, Ubuntu 20.04 or newer on Linux, and AVX2 CPU support on x64 — that last one rules out some older machines whatever GPU you pair with them. It recommends 16GB of system RAM and at least 4GB of dedicated VRAM, though 8GB is the practical floor for 7B models with usable context. LM Studio also displays whether your GPU has enough VRAM before loading a model, so you can check a specific model before downloading it.
Does LM Studio support multiple GPUs?
LM Studio can split model layers across multiple NVIDIA GPUs when a single card does not have enough VRAM. However, multi-GPU support is not as seamless as single-GPU use — you may need to manually configure layer allocation, and inter-GPU communication adds some overhead. For most users, a single high-VRAM card like the RTX 4090 is simpler and often faster than two smaller cards.
How much VRAM does LM Studio need?
VRAM needs depend on the model size and quantization level. For 7B models at Q4, you need about 5GB. For 13B models at Q4, about 8GB. For 34B models at Q4, about 20GB. LM Studio also uses VRAM for the KV cache during conversations, so budget an extra 2-4GB beyond the base model size for comfortable context lengths.
Does LM Studio work on Apple Silicon with MLX?
Yes, and it is one of LM Studio’s biggest advantages. LM Studio automatically selects the MLX backend on Apple Silicon Macs, which uses unified memory efficiently. An M5 Pro with 24GB handles 13B models well, and an M5 Max with 48GB runs 34B models comfortably. MLX performance on Apple Silicon often matches or exceeds mid-range NVIDIA GPUs for equivalent model sizes.