dalao.ai

LM Studio

Find, download and run local LLMs with a desktop GUI

About LM Studio

LM Studio, from Element Labs, is a desktop app for downloading and running open-weight models on Mac, Windows and Linux through a GUI. GGUF models run on llama.cpp; on Apple Silicon, MLX models can run on Apple's MLX. It is also a local API server at http://localhost:1234 with OpenAI- and Anthropic-compatible endpoints. The app is closed source and free for personal and internal business use; the lms CLI, SDKs and MLX engine are MIT. In 2026 the company also launched Bionic, a separate agent app, and offers paid US-hosted cloud inference.

EditorialUpdated 12 min read

The LM Studio interface

What it is

The basic loop takes a few clicks. Search the Discover tab, download a model, load it and chat; or open the Developer tab and serve the loaded model as a local API. No engine compiling or terminal work is needed. The docs say that once models are on disk, the app works fully offline.

System requirements (as of 2026-10-25). The app docs summarize support as Apple Silicon Macs, x64/ARM64 Windows PCs and x64 Linux PCs; the system requirements page goes further. On macOS it needs Apple Silicon (M1–M4) and macOS 14.0 or newer; 16GB of RAM is recommended, and Intel Macs are not supported. Windows builds cover x64 and ARM (Snapdragon X Elite); x64 CPUs need AVX2, and the docs suggest 16GB of RAM and at least 4GB of dedicated VRAM. Linux builds cover x64 and ARM64 and ship as an AppImage for Ubuntu 20.04 or newer. The docs warn that releases after 22 are less tested.

Current version. As of 2026-10-25 the latest release was 0.4.26. Its changelog header reads October 13, 2026 · Build 5. Individual builds carry no dates of their own; as of 2026-10-25 the latest was Build 5.

Company. The developer is Element Labs. Its SEC Form D, filed May 30, 2025, gives 2022 as the year of incorporation and Modifiable Inc. as a former name. It lists Yagil Burowski as executive officer, director and promoter, and Matt Turck as a director. The filing reports $19,324,955 of equity sold, with the first sale on April 25, 2025, to 16 investors. That figure includes $3.965 million from converted convertible securities. The form does not name investors. A July 8, 2025 company blog post put the team at nine people plus two summer interns.

Where Bionic fits. On July 16, 2026 the company launched LM Studio Bionic, which it describes as "a new, separate app from LM Studio", built as an agent for work and coding. The launch post says local models in Bionic "are powered by the LM Studio runtime". Since 0.4.26 Build 3 the two apps can run side by side, with Bionic using LM Studio for local inference (Bionic 1.1.8 or later required).

Starting today, there's no need to fill a form or contact us. You and your team can just use LM Studio at work!

LM Studio blog, July 8, 2025

What stands out

Two engines behind one UI. Per the docs, LM Studio runs models with llama.cpp on all three desktop platforms, and on Apple Silicon Macs it can also use Apple's MLX. The llama.cpp side ships in several variants: CPU-only, CUDA, Vulkan, ROCm and Metal. Each engine reads its own format: GGUF for llama.cpp, MLX weights for MLX. For CLI downloads, lms get --gguf or lms get --mlx limits results to that format (without a flag, results are filtered by the runtimes present).

Engines update apart from the app. LM Studio packages its engines as separate libraries called LM Runtimes. Since 0.3.0 new runtimes can be downloaded without a full app update, and lms runtime select switches the active one. They are managed in two places. In the app, ⌘⇧R (Mac) or Ctrl+Shift+R (Windows/Linux) opens the runtimes page. In a terminal, lms runtime has five subcommands: ls lists runtimes, get downloads one, select sets the active one, remove uninstalls one and update upgrades one. Examples are lms runtime update llama.cpp and, on macOS, lms runtime update mlx.

Its own llama.cpp version numbers. LM Studio numbers its llama.cpp runtime on its own scheme and notes the upstream build in the changelog. 0.4.26 Build 2 reads "Updated the llama.cpp runtime to 2.50.0 (upstream b11337)". GitHub shows b11337 published on 2026-10-02; by 2026-10-25 the newest upstream build tag was b11871 (marked prerelease on GitHub), and the latest tagged release was v0.7.0 (2026-10-16). Inference: upstream llama.cpp features reach LM Studio users after some delay, and how long depends on the company's packaging cadence.

The MLX engine is open. mlx-engine is MIT-licensed. Its README lists three open-source building blocks: Apple's mlx-lm for text, mlx-vlm for vision and Outlines for structured output. The README also says every Mac build since 0.3.4 bundles it. On 2026-10-25 the repo had 1,194 stars, and its last push was on 2026-10-21.

Not just a desktop app. Version 0.4.0 (January 28, 2026) split the GUI from the core and shipped the core as llmster, a daemon for headless Linux boxes, cloud servers and CI. The same release added parallel requests to one model through continuous batching instead of a queue. The company credits llama.cpp's open-source implementation for this, and at launch the feature was not yet in the MLX engine ("actively in the works").

LM Link across devices. The docs describe LM Link as an end-to-end-encrypted network built with Tailscale. It allows running a big model on one machine and using it from a laptop or an iPhone (through the Locally app). The Free tier on the Bionic pricing page lists LM Link for up to five devices (as of 2026-10-25).

How it really differs from Ollama

Interface. LM Studio is built around its GUI, with lms and headless llmster as add-ons. Ollama is mainly a CLI and API, but it also has desktop apps: its GitHub README offers a macOS .dmg and a Windows .exe.

License. The LM Studio app runs under Element Labs' terms of service (version of August 23, 2026), which grant a non-exclusive, non-transferable license to use the software. That is not an open-source license. Its lms CLI, SDKs and mlx-engine are MIT. Ollama's main repo (local runtime and clients) is MIT-licensed (as of 2026-10-25); its cloud service is not part of that repo.

API compatibility. By default LM Studio serves on port 1234. It exposes OpenAI-compatible routes (/v1/models, /v1/responses, /v1/chat/completions, /v1/embeddings, /v1/completions and more), an Anthropic-compatible /v1/messages and its own REST API. Ollama listens on 11434 by default and also documents OpenAI- and Anthropic-compatible endpoints. With either one, existing OpenAI or Anthropic clients only need a new base URL.

Model formats. LM Studio runs GGUF on llama.cpp and, on Apple Silicon, MLX weights on MLX.

Cloud. LM Studio's pricing page (as of 2026-10-25) lists Free at $0 for local models, plus Bionic+ at $20/month and Pro at $100/month for US-hosted open-model inference. Ollama's pricing page (as of 2026-10-25) is likewise free for local use. Its cloud tiers are Pro at $20/month ($16.67/month billed yearly), Max at $100/month and an early-access Team plan at $500/month.

Server deployment. Since 0.4.0, LM Studio's headless option has been llmster, set up by a one-line script; an official guide shows how to run it as a systemd service on Linux. As of 2026-10-25 its docs offer no official Docker image. Ollama's README points to the official ollama/ollama Docker image.

In short. LM Studio centers on a GUI and its own Bionic agent. Ollama is mainly a CLI and API, and its local parts are open source.

Who it's for, and who it isn't

Who it's for

People who want to try open models on their own computer without a terminal. Search, download, load and chat all happen in one window.

Apple Silicon owners. The same app offers both GGUF (llama.cpp) and MLX builds of a model, with no Python setup.

Developers who need a local OpenAI- or Anthropic-compatible endpoint for tools like Codex or Claude Code. The docs show how to wire them up.

Teams that want a cheap first look at local models at work. Since July 2025, workplace use no longer needs a separate commercial license.

Not for

Teams that need an open, auditable stack they can modify. The app is closed, and its terms prohibit modification, reverse engineering and redistribution.

Anyone planning to resell inference as a service or bundle the app with a product. The terms bar sublicensing, distribution, sale and "service bureau" or SaaS use.

Intel Mac users and owners of Windows x64 PCs without AVX2. Neither is supported.

Ops teams that expect a standard Docker workflow. As of 2026-10-31 the official docs offer no Docker image; deployment is an install script plus a daemon.

Getting started

Desktop. Install the app from the download page (an AppImage on Linux). In Discover, search for a small model that fits the machine's RAM, download it, select it in the top bar and start chatting. To change engine versions, open the runtimes page with ⌘⇧R or Ctrl+Shift+R.

Local server. Start the server from the Developer tab or run lms server start; it listens on http://localhost:1234 by default. With the OpenAI SDK, set base_url to http://localhost:1234/v1. With the Anthropic SDK, set it to http://localhost:1234.

Headless (llmster). Run curl -fsSL https://lmstudio.ai/install.sh | bash on Linux or Mac, or irm https://lmstudio.ai/install.ps1 | iex on Windows. Then lms daemon up starts the daemon, lms get <model> fetches a model and lms server start serves it. The official Linux guide checks the llmster install with lms --help and provides a systemd unit that loads a model and starts the server at boot.

From code. The official lmstudio-js and lmstudio-python SDKs, both MIT, can load models, chat and call tools from other programs.

Pricing & license

Personal use. Free. The terms (version of August 23, 2026) grant a non-exclusive, non-transferable license to use the software "solely for Your personal and / or internal business purposes".

Use at work. Also free. A July 8, 2025 post says a separate commercial license is no longer needed for work use, so there is no form to fill in and no one to contact. When an employee uses the app for their job or mainly for the company's benefit, the terms treat the company itself as the customer. An Enterprise plan adds SSO, model and MCP gating and private collaboration, per that post. Pricing is through sales.

Where commercial use stops. Internal business use is allowed. The terms forbid modifying the software or making derivative works, integrating it other than through published interfaces, using it with open-source products in a way that could require the software to be disclosed in source form, licensed for modification or redistributed at no charge (inference: this mainly targets copyleft licenses such as the GPL), and sublicensing, distributing, selling, leasing, renting or lending it. They also forbid using it "for service bureau use, as an application service provider, or a software-as-a-service". The source code is treated as a trade secret, and reverse engineering is off-limits unless local law expressly permits it. Inference: the terms don't say in so many words whether a paid external product that gets its inference from LM Studio's local server counts as SaaS or service-bureau use. The terms do not define which uses count as SaaS or service-bureau use.

Paid tiers (as of 2026-10-25). The pricing page now lists Bionic plans only:

  • Free, $0: the Bionic agent, local models on llama.cpp and MLX, offline voice transcription, LM Link on up to five devices and limited web search.
  • Bionic+, $20/month: adds US-hosted open models (the page names Kimi K3, GLM 5.3 and DeepSeek V4 Flash, among others), discounted bulk tokens, web search and page extraction.
  • Pro, $100/month: everything in Bionic+ with "5× usage limits", plus early access to new features.

Cloud credits pay for usage beyond the plan allowance. Organizations get centralized billing for credits, and team subscriptions are marked "coming soon". Under the terms, subscriptions renew automatically, a cancellation takes effect at the end of the billing period, fees are non-refundable unless the law says otherwise, allowances don't roll over, and the company may change features, limits and prices.

Open-source parts (repo LICENSE files, as of 2026-10-25).

  • lms CLI: MIT, "Copyright (c) 2024 LM Studio", 5,379 stars.
  • lmstudio-js: MIT, "Copyright (c) 2025 Element Labs Inc", 1,802 stars.
  • lmstudio-python: MIT, "Copyright (c) 2025 LM Studio", 886 stars; latest stable release 1.5.0 (2025-08-22), followed only by the 1.6.0b1 beta (2025-10-18); last commit 2025-10-27.
  • mlx-engine: MIT, "Copyright (c) 2024 LM Studio", 1,194 stars.

None of the public lmstudio-ai repos contains the llama.cpp wrapper, which the 0.4.0 post calls LM Studio's llm-engine. Inference: that layer is closed, like the app.

Risks & limits

Closed source, changeable terms. The app is proprietary. The company can post new terms, and continued use counts as acceptance; it can also add, cut or discontinue features. Liability is capped at $50, and New York law governs. The terms of service restrict use in some regions and for some purposes; see the terms for details.

No auth by default. The docs say the API runs without enforcing authentication out of the box; it is enabled with "Require Authentication" in the server settings. Once it is on, tokens with granular permissions can be issued, sent as either x-api-key or Authorization: Bearer. "Serve on Local Network", or lms server start --bind 0.0.0.0, exposes the server to the LAN, and the docs recommend enabling auth in that case.

Security advisories (as of 2026-10-25). As of 2026-10-25, NVD lists no CVE with LM Studio as the affected product. The lms, lmstudio-js, lmstudio-python, mlx-engine and bug-tracker repos publish no GitHub security advisories. As of 2026-10-25, the GitHub advisory GHSA-g5f9-3xfg-p9mf (CVE-2026-61732) mentions LM Studio in its text, but it targets a different project, Decepticon (versions before 1.1.17):

  • GitHub rates it Critical.
  • The advisory title is "Role-boundary forgery via ChatML special-token literals in web crawl output composed into LLM context".
  • The CVSS 3.1 score is 10.0.
  • NVD lists it as Deferred with no score of its own; the 10.0 shown there comes from GitHub.

The advisory says most self-hosted model servers, LM Studio among them, don't filter special-token literals from user content by default. That is a third-party description, and the advisory did not test LM Studio itself: its applicability section says "vLLM / SGLang / TGI confirmed upstream". As of 2026-10-25 LM Studio has not published a response.

Platform limits. There is no Intel Mac support, Windows x64 needs AVX2, and Linux ships as an AppImage documented for Ubuntu 20.04+. The app docs list Linux as x64, while the system requirements page also lists ARM64; the system requirements page is the more detailed source.

No Python SDK commits since 2025-10-27. lmstudio-python's latest stable release is still 1.5.0 (2025-08-22), followed only by the 1.6.0b1 beta (2025-10-18), and its last commit was on 2025-10-27 (all as of 2026-10-25). Over the same period lmstudio-js kept moving, with a push on 2026-10-24. Inference: new APIs may show up first in the JS SDK or the REST API rather than the Python SDK.

Engines aren't at parity. At the 0.4.0 launch (January 28, 2026), parallel requests worked only on the llama.cpp engine, with MLX support still in progress. The same model can offer different features in its two formats.

A shifting focus. The pricing page lists only Bionic plans, the Bionic launch post promises zero data retention (ZDR) for cloud inference, and the pricing page says the cloud models are US-hosted. Inference: new features may land in Bionic before the LM Studio app itself.

Our verdict

LM Studio suits running open models locally with minimal setup, especially on Apple Silicon. It offers a GUI, two engines, runtimes that update apart from the app and a local server that speaks both the OpenAI and Anthropic APIs, which covers most individual and small-team needs. And since July 2025, using it at work costs nothing and needs no paperwork.

The price is that it's closed. Neither the app nor its llama.cpp layer is public, the terms rule out modification, redistribution and offering it as a service, and security review means trusting the vendor. For teams that need an open-source local stack, a standard Docker deployment or inference built into their own product, Ollama or llama.cpp itself is a better fit.

Suitable use: personal and internal experiments. The local server does not enforce authentication by default, and the terms restrict external services and redistribution.

Why it stands out

LM Studio is a desktop app for Mac, Windows and Linux that finds, downloads and runs open-weight models through a GUI. GGUF models run on llama.cpp; on Apple Silicon it can also run MLX models on Apple's MLX. With a model loaded it doubles as a local API server at localhost:1234, with OpenAI- and Anthropic-compatible endpoints.

It is made by Element Labs, Inc., whose 2025 SEC filing lists Yagil Burowski as an executive and director. The desktop app is closed source and, since July 2025, free for personal and internal business use; the lms CLI, the JS and Python SDKs and the MLX engine are MIT licensed. In July 2026 the company launched Bionic, a separate agent app that runs on the LM Studio runtime.

It runs open-weight models through a GUI on Mac, Windows and Linux, including MLX models on Apple Silicon, and doubles as a local API server with OpenAI- and Anthropic-compatible endpoints.

Sources

Official

  1. LM Studio App Terms of Servicelmstudio.ai · Accessed Oct 25, 2026
  2. LM Studio is free for use at worklmstudio.ai · Accessed Oct 25, 2026
  3. Bionic Pricing | LM Studiolmstudio.ai · Accessed Oct 25, 2026
  4. Introducing LM Studio Bionic: the AI agent for open modelslmstudio.ai · Accessed Oct 25, 2026
  5. Changelog • LM Studio 0.4.26lmstudio.ai · Accessed Oct 25, 2026
  6. LM Studio 0.3.0lmstudio.ai · Accessed Oct 25, 2026
  7. LM Studio 0.3.9lmstudio.ai · Accessed Oct 25, 2026
  8. Introducing LM Studio 0.4.0lmstudio.ai · Accessed Oct 25, 2026