Ollama vs LM Studio: the two garages your models live in

■ space · the two garages

THE TOUR

Say what you came for. The garages will answer.

THE PICK: Ollama — the service garage. It starts at boot and answers on :11434 while your box stays mouse-free.

# the whole handshake
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:7b
curl http://localhost:11434/api/chat -d '{
  "model": "qwen2.5-coder:7b",
  "messages": [{"role":"user","content":"Explain bash exit codes in two lines"}],
  "stream": false
}'

# Developer tab → start server — or the lms CLI
lms server start
lms load qwen2.5-coder-7b-instruct --gpu max
curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen2.5-coder-7b-instruct","messages":[{"role":"user","content":"Write a jq filter for the top process by CPU"}]}'

the licence drawer: Ollama is MIT — you only owe the model's licence (Llama, Mistral, Qwen and Gemma each carry their own terms). LM Studio is free for personal use; work use needs its free work-licence flag, and companies above a revenue threshold pay. Procurement asks — this row ends the debate.

WHAT THE PROCESSOR COLUMN IS TELLING YOU

Ollama detects CUDA, Metal and ROCm at launch and offloads everything that fits. The check is one command — the answer is one word. Flip it:

# what did Ollama actually load — GPU or CPU?
ollama ps

# force the issue if it silently fell back to CPU:
OLLAMA_NUM_GPU=999 ollama run qwen2.5-coder:7b

OLLAMA_NUM_GPU overrides the layer count — big number: everything on the card; 0: CPU only. A probe — the real fix is a quant that fits.

LM Studio exposes the same control as a GPU offload slider per model — friendlier to experiment with, one of its genuinely better touches.

THE SPEED QUESTION, SETTLED

Same model, quantisation, hardware: effectively a tie — both delegate to llama.cpp. Claims of a gap usually compare different quants or context lengths. Measure on your own machine:

# Ollama: --verbose prints eval rate (tokens/sec) at the end
ollama run qwen2.5-coder:7b --verbose "Summarise what a Makefile does in one sentence."

In LM Studio, load the same GGUF with identical context and GPU layers; watch tokens/sec in the stats panel. Rerun twice — first load is warm-up noise.

FIND YOUR ROW

Rules of thumb, no regrets either way:

1. You want to use models — chat, try them out, fiddle with sliders: LM Studio.
2. You want to build on models — scripts, agents, CI, a home API service: Ollama.
3. You want both — install both; they coexist fine. Just don't let both servers claim the same port, and remember a model loaded twice eats VRAM twice.

Both understand GGUF and drive the same lineage — Ollama embeds llama.cpp; LM Studio uses llama.cpp-based runtimes. Same file, same output; the tools differ around the model.

Local models pair especially well with agentic coding tools: point an OpenAI-compatible client at your endpoint — unlimited completions at zero per-token cost. The only bill is the electricity.

end of the two garages · nothing fetched · everything local