olace

[ local ai ]

LM Studio

LM Studio has first-class Apple Silicon support through MLX, which is why Olace recommends it on Macs.

LM Studio, llmster, and lms

The name covers three separate things, and Olace only needs one of them:

  • LM Studio is the desktop app, the graphical interface.
  • llmster is the headless daemon: the same engine, packaged to run as a background service with no GUI. This is what actually serves your models.
  • lms is the command line tool, and it drives either one.

Olace installs and talks to llmster. You do not need the LM Studio app, and Olace never opens it. If you do run it, Olace uses the llmster inside it just the same. Quitting the LM Studio app stops that llmster, but Olace starts it back up headlessly, right away when a message needs it and within a couple of minutes otherwise, so the machine stays reachable either way. LM Studio's own explanation of the three names goes into more detail.

Install

Easiest: Settings › Local AI in the Olace app, or olace setup --provider lms. Manual install:

# Linux and macOS
curl -fsSL https://lmstudio.ai/install.sh | bash

# Windows (PowerShell)
powershell -NoProfile -Command "irm https://lmstudio.ai/install.ps1 | iex"

That is LM Studio's own llmster installer, and it needs no GUI and no root. It puts the lms CLI in ~/.lmstudio/bin; lms daemon up starts the daemon, though Olace does that for you on demand. Olace talks to it at http://localhost:1234 (configurable via lmstudio_url in daemon.json).

Models

LM Studio model ids use a slash, like google/gemma-4-e4b or mistralai/ministral-3-3b. Browse the catalog at lmstudio.ai/models, or use the Model hub in the Olace app, which curates it and labels what fits your hardware. From the terminal:

olace pull google/gemma-4-e4b --provider lms
olace list
olace pull google/gemma-4-e4b --device home-pc   # install on a paired host, remotely

Managed for you

  • Starts on demand. The daemon starts llmster when a request needs it. You do not have to keep anything running.
  • Keep-alive and unload work the same as with Ollama: olace settings keep-alive, olace unload.
  • Optimization. Olace detects and configures Flash Attention and KV cache settings per model load where llmster supports them.
  • Context sizing. Every model gets its own context window, sized from what the model costs and what your GPU can spend, then tuned as Olace learns what actually runs on your machine. See Hardware & context.

Running more than one runtime

LM Studio, Ollama, and llama.cpp coexist cleanly. Models from all of them show up in one library, and each runtime's context sizing is tuned independently.