Installation and first model: under 5 minutes
Download the installer from lmstudio.ai (Windows, macOS, or Linux AppImage). Launch it, and the home screen has a search bar for HuggingFace models. Type 'deepseek', see 50+ GGUF versions of DeepSeek models, pick one with a green compatibility badge, click download. The 8.5GB Q4_K_M version of DeepSeek-R1 14B downloaded in 4 minutes on my connection. Click 'Load Model' and you have a working chat interface. The entire flow takes less time than signing up for ChatGPT Plus.
The model browser solves the biggest local AI problem
Finding the right GGUF model on HuggingFace is a mess: 50,000 files, cryptic quantization names, no indication of which ones work with your hardware. LM Studio's model browser shows only compatible models, highlights the recommended quantization, displays RAM requirements, and shows download counts. The green badge means 'this model will run on your hardware.' This alone makes LM Studio worth installing over manually downloading GGUF files and figuring out the right quantization.
The local server feature: why I keep LM Studio running
LM Studio has a one-click toggle to start an OpenAI-compatible API server at `localhost:1234`. I turn this on and point Continue.dev (my VS Code AI plugin) to `http://localhost:1234/v1`. Now I have code completion powered by a local DeepSeek model with zero latency and zero API cost. The server UI shows request history, response times, and GPU memory usage in real time. For development work where I do 200+ completions per day, this saves $3-5 daily in API costs.
LM Studio vs Ollama: when to use which
LM Studio is a desktop app with a GUI. Ollama is a CLI tool with a REST API. If you want to double-click and chat, use LM Studio. If you are building a pipeline that needs to call models from code, use Ollama. LM Studio's model download and management is better (visual browser, compatibility badges, RAM estimates). Ollama's server is more battle-tested (lower latency, better concurrent request handling). I use LM Studio for exploration and prototyping, Ollama for production workflows.
Hardware requirements and real performance
On my RTX 3090 (24GB): DeepSeek-R1 14B Q4 runs at 55 tokens/sec (faster than reading speed). Llama 3.3 8B Q8 runs at 80 tokens/sec. On my MacBook M2 with 16GB RAM: DeepSeek 7B Q4 runs at 18 tokens/sec (reads slightly faster than typing). On CPU-only laptop with 8GB RAM: Gemma 2B Q4 runs at 12 tokens/sec (usable for simple Q&A). LM Studio shows real-time GPU/CPU utilization and lets you offload specific numbers of layers to GPU. The slider UI for GPU offloading is the clearest implementation I have seen.
