DEV Community

Jovan Chan
Jovan Chan

Posted on • Originally published at aifoss.dev

LM Studio Review 2026: Easiest Way to Run Local LLMs on Mac and Windows (0.4 Tested)

#lmstudio #ai #llm #selfhosted

This article was originally published on aifoss.dev

LM Studio has become the go-to answer for anyone who wants to run local LLMs but doesn't want to touch a terminal. Version 0.4.13 ships with a polished model browser, a built-in chat interface, and a local OpenAI-compatible API — all wrapped in a desktop app that actually feels finished.

The catch: it's not open source. The core app is proprietary, free for personal and commercial use as of late 2025. If that's a dealbreaker, stop here and use Ollama. If you can live with a closed binary and want the easiest path to running Qwen 3 8B or Gemma 4 on your laptop, LM Studio is hard to beat.

The one-paragraph verdict: LM Studio 0.4.13 is the best GUI option for local LLMs right now, especially on Apple Silicon where its MLX backend significantly outpaces Ollama's GGUF path. It falls short for server deployments, scripted automation, and any context where you need auditable source code.

What LM Studio is (and isn't)

LM Studio is a desktop application — available for macOS, Windows, and Linux — that downloads GGUF-format models from Hugging Face, handles configuration through a GUI, and runs a local API server compatible with the OpenAI SDK.

What it is not:

  • A model itself — it's a runtime and interface, not weights
  • An open-source tool — the core app is closed-source (the lms CLI companion has an MIT-licensed repo, but the main application does not)
  • A good choice for headless servers — Docker support for GPU is in preview and currently CPU-only on x86

The design philosophy is clear: minimize friction for people who know what they want to do but don't want to manage daemons, flags, and YAML. If you've used Ollama and found the terminal overhead annoying, LM Studio is the correction.

License: Proprietary, closed-source. Free for personal and commercial use. No license request, no form — just download and run. If your threat model includes binary auditing, this is a real constraint.

Installation

Download the installer from lmstudio.ai. No package manager required, no dependency resolution. On macOS it's a standard DMG. On Windows, a standard installer. The Linux version (AppImage) has been in active development since late 2024 and is stable enough for daily use.

First-launch experience: you land in a model browser that queries Hugging Face in real time. Search for "Qwen3 8B," click download, and it handles the rest — including checksum verification.

Minimum hardware per LM Studio's official system requirements:

  • RAM: 16 GB recommended (8 GB workable for 3–4B models only)
  • VRAM: 4 GB dedicated minimum; 8 GB for comfortable 7B use
  • CPU fallback: any x86_64 or ARM64 CPU if no GPU is available, just slow

For Apple Silicon Macs, the unified memory architecture is a real advantage. The MLX backend LM Studio uses on Apple Silicon shares memory between CPU and GPU, so a 16 GB M3 Pro can run 7B models comfortably that would require a discrete 8 GB VRAM card on a PC. If you're looking at GPU upgrades for local LLM work on PC, an RTX 4070 or 4080 hits the sweet spot for 13–30B model use.

The model browser

This is where LM Studio earns its reputation. The browser pulls directly from Hugging Face and shows you file size, quantization level (Q4_K_M, Q5_K_M, Q8_0, etc.), and estimated VRAM usage before you download. Most competing tools make you find and paste a model URL manually.

Quantization guidance is shown inline — a 7B model at Q4_K_M needs roughly 4.5 GB of VRAM, the same model at Q8_0 requires about 8 GB with better output quality. This pre-download information is something Ollama's CLI doesn't surface without digging through documentation.

You can also load models from a local path, which matters if you're working with fine-tuned models or anything not on Hugging Face.

Chat interface

The built-in chat UI is competitive with Open WebUI without needing a separate server running. Multi-turn conversations, system prompt configuration, and parameter sliders (temperature, top-p, context length) are all accessible from the main window.

Version 0.4.13 includes PDF chat — load a PDF directly into the context. It's basic retrieval (not indexed), but functional for single-document Q&A without setting up a full RAG pipeline.

One thing that distinguishes LM Studio from Open WebUI: the model parameter controls are visible per-conversation rather than buried in settings. If you're experimenting with temperature settings across multiple runs, that's a meaningful difference in workflow.

The local API server

Start the API server from the "Local Server" tab. Default port: 1234. Endpoint: http://localhost:1234/v1.

Any code written for the OpenAI Python SDK works immediately by changing the base URL:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio"  # arbitrary string, not validated locally
)

response = client.chat.completions.create(
    model="qwen3-8b-q4_k_m",
    messages=[{"role": "user", "content": "Explain GGUF quantization in one paragraph."}]
)
print(response.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

The server also supports the embeddings endpoint, which means you can use it as the backend for RAG workflows without changing your application code.

LM Studio 0.4.12 added MCP OAuth support, making it possible to connect external tools — file servers, web fetchers, code execution environments — through the Model Context Protocol. This brings it closer to the agentic use cases that previously required a more complex setup.

LM Link, introduced in early 2026, extends this further: expose a remote LM Studio instance over an encrypted connection and use it as if it were local. Useful for running a beefy desktop machine headlessly while working from a laptop.

Performance: Apple Silicon vs. Windows/Linux

Platform matters significantly here.

On Apple Silicon (M2/M3/M4), LM Studio's MLX backend outperforms Ollama's GGUF path in published 2026 benchmarks. On M3 Ultra hardware, Gemma 3 1B reached 237 tok/s in LM Studio versus 149 tok/s in Ollama — a roughly 59% difference attributable to the MLX engine's use of Apple's unified memory. If you're on an Apple Silicon Mac, this is a genuine reason to prefer LM Studio, not a marketing claim.

On Windows and Linux with NVIDIA GPUs, the picture reverses. Ollama's inference overhead is lower — roughly 100 MB of process memory versus LM Studio's ~500 MB GUI footprint — and it runs 10–20% faster in inference-only scenarios. For a server running 24/7 without a user at a GUI, that overhead compounds.

LM Studio vs. the alternatives

Feature LM Studio 0.4.13 Ollama Jan.ai
Interface GUI desktop app CLI + API daemon GUI desktop app
License Proprietary, free MIT (open source) AGPL-3.0 (open source)
Apple Silicon backend MLX (fast) GGUF (slower) GGUF
NVIDIA GPU support ✓ CUDA ✓ CUDA ✓ CUDA
Memory overhead ~500 MB ~100 MB ~300 MB
Docker / headless Preview (CPU-only) Full GPU passthrough Limited
Model browser Built-in, Hugging Face CLI ollama pull Built-in, limited
OpenAI-compatible API ✓ localhost:1234 ✓ localhost:11434 ✓ localhost:1337
MCP support ✓ v0.4.12+ ✓ Partial
Source auditable ❌ ✓ ✓

vs. Ollama: Ollama is the right choice for server deployments, automation pipelines, Docker, and any context requiring auditable source code. LM Studio wins for first-time setup, model discovery, and Apple Silicon performance. The practical recommendation: use LM Studio to find and evaluate models, then switch to Ollama to operationalize them in production.

vs. Jan.ai: Jan.ai is fully open source (AGPL-3.0) with a similar desktop-first design. The interface is less polished and the model library is smaller, but it's the FOSS alternative if the proprietary nature of LM Studio is a hard constraint.

**vs. l

Top comments (0)