LM Studio

tl;dr: LM Studio is no longer just a chat GUI. The current build — Bionic 1.1.x — is an agent that writes documents, slides, PDFs, and code using local models (MLX + llama.cpp) or frontier open models via a zero-data-retention cloud.

What it does

Bionic is LM Studio’s agent layer: it creates and edits documents with autosave, runs coding tasks, automations, and computer control. Underneath sits the LM Studio runtime with MLX and llama.cpp engines, real-time local voice transcription (audio stays on device), and an in-app model downloader.

Latest release is Bionic 1.1.2 (Sep 8, 2026): UI performance work, llama.cpp 2.33.0 extension packs, more reliable MLX model detection, drag-and-drop sessions, and the first Linux builds (x64 and ARM64).

Installation

Download from lmstudio.ai for macOS, Windows, or Linux. The lms CLI is bundled; 1.1.2 added lms load --auto for scripted model loading.

Agent features

• Documents, slides, and PDFs with tracked changes and comments (Word), created in sessions
• Coding projects with shell-command Auto Review and persistent approval preferences
• Skills: built-in plus custom, with /install-skill and /create-skill commands
• MCP servers, including OAuth-authenticated ones
• Vision subagents: hand image understanding to a second model when your main model is text-only
• Sandboxed Pyodide operations for crypto and other isolated tasks

Local runtime

Download GGUF/MLX models in-app and run them fully offline. 1.1.0 brought 2–2.75× faster prompt processing on M5 Macs, MLX prompt disk cache controls, and Muse Glimmer tool calling for MLX models. llama.cpp extension packs track upstream closely (2.33.0 as of 1.1.2).

Local API server

Developer mode still serves an OpenAI-compatible API. As of 1.1.0, /v1/responses, /v1/chat/completions, and /v1/messages all accept images in tool call results.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")
response = client.chat.completions.create(
    model="local-model",  # whatever is loaded in LM Studio
    messages=[{"role": "user", "content": "Hello!"}]
)

Cloud models

For heavy tasks, Bionic can run frontier open models — GLM 5.2, Kimi K3, DeepSeek V4 Pro — through LM Studio’s cloud. All cloud services are Zero Data Retention. Cloud plans and credit usage are managed in-app.

vs Ollama

Gotchas

• The lmstudio-ai/lmstudio-js GitHub repo is the SDK, not the app — its star count is not LM Studio’s popularity, so we don’t cite one.
• Proprietary and closed-source. If that matters, use Ollama.
• Linux support only landed in Bionic 1.1.2 — expect rougher edges than macOS.

Related: Ollama for the MIT CLI runtime, Open WebUI if you want a browser UI on this API, best local LLM, Ollama vs LM Studio.

Untested here — facts from lmstudio.ai and the Bionic changelog fetched 2026-09-10.