LM Studio is a desktop app for downloading AI models and running them on your own computer, with a chat interface and a local server you can connect other tools to. It’s made by Element Labs, it’s free to use at work too as of July 8, 2025, and the current version is 0.4.28, released October 31, 2026. I can set up local AI for you if you don’t want your data leaving the building, and train the team that will use it. Here’s what it is, what its license actually lets you do, what changed with Bionic and when something else is a better fit.
What LM Studio is (and what it isn’t)
The download page sums it up in one line: “Experiment with LLMs on your computer. Chat interface and programmable API.” According to its documentation, it comes with a model search and downloader that pulls from Hugging Face, a chat where you can attach documents and query them offline (RAG), and a local server that speaks the same language as the OpenAI API.
Three things most search results don’t make clear:
- It isn’t made by Hugging Face. It downloads models from there, but the company behind it is Element Labs, Inc., the name in the site footer and on the Terms of Use (which are governed by New York law). In July 2025, its blog said the team was 9 people plus 2 summer interns. The app launched in May 2023.
- It isn’t open source. It’s free, but the app is proprietary: the Terms prohibit reverse engineering, decompiling or disassembling it. What is open is the
lmsCLI, the TypeScript and Python SDKs and themlx-engine(all three under the MIT license). - It isn’t Bionic. Today, the lmstudio.ai homepage promotes a different app, LM Studio Bionic (launched in July 2026), an agent with a paid cloud. LM Studio is still around and still getting updates; I explain the difference below.
Watch out for third-party apps too: Google Play lists an app called “LM Studio Local AI Guide” that isn’t official.
Is LM Studio free? Commercial use and licensing
Yes, including at work. On July 8, 2025, Element Labs announced it on its blog: “Starting today, LM Studio is free to use both at home and at work.” Until then, using it inside a company meant requesting a separate commercial license; the same post puts it plainly: “This requirement is now removed.” And it adds: “You and your team can just use LM Studio at work!”
Here’s what the Terms of Use say today, in the version dated August 23, 2026:
| Point | Official wording | What it means |
|---|---|---|
| What the license covers | “solely for Your personal and / or internal business purposes” | Personal use or internal use within your company |
| What you can’t do | “sublicense, distribute, sell, use for service bureau use, as an application service provider, or a software-as-a-service” | You can’t resell LM Studio or offer it as a service to third parties |
| If you use it as an employee | The Terms are between your company and Element Labs, not you | Your company is the one accepting the terms |
| Models | Third-party models come with “a license agreement between you and the third party provider” | Every model has its own license, separate from LM Studio’s |
| Paid features | Some features are charged by subscription, by usage or with credits (“Paid Features”) | Today, that’s the Bionic cloud |
| Liability | Capped at “$50.00” in total | Barring exceptions such as fraud, Element Labs is liable for no more than 50 dollars |
◆ What the license doesn’t spell out. “Internal use” rules out reselling the app, but the Terms don’t explain how it works when a consultant installs it on a client’s machines. I couldn’t confirm that. My advice: have the company itself accept the Terms for its own internal use and, if you’re going to build a service for third parties on top of it, check with Element Labs or your lawyer first. And read the license of every model you download: they don’t all allow the same things.
Version, platforms and system requirements
The current version of LM Studio is 0.4.28, released October 31, 2026, and published “under the terms of use.” The big leap was 0.4.0 (January 28, 2026), which brought, according to its blog, “Server deployment, parallel requests with continuous batching, new REST API endpoint, and refreshed application UI”: server deployment and handling several requests at once, which brings it closer to a lightweight server.
Requirements from the official documentation, checked on November 16, 2026:
| OS | Processors | Requirements and recommendations |
|---|---|---|
| macOS | Apple Silicon (M1, M2, M3, M4). Intel Macs aren’t supported: “Intel-based Macs are currently not supported” | macOS 14.0 or later. 16 GB of RAM recommended; with 8 GB it works if you stick to small models and modest context sizes |
| Windows | x64 and ARM (Snapdragon X Elite) | On x64, the processor must support AVX2. At least 16 GB of RAM and 4 GB of dedicated VRAM recommended |
| Linux | x64 and ARM64 (aarch64) | The docs say it’s distributed as an AppImage; the download page also offers a .deb package. Ubuntu 20.04 or later; versions above 22 are “not well tested” |
Once you’ve downloaded your models, it doesn’t need an internet connection: “LM Studio can operate entirely offline, just make sure to get some model files first.”
Engines, formats and models
Under the hood, LM Studio uses two engines: llama.cpp and MLX, the latter for Apple Silicon Macs. The homepage, which now showcases Bionic, sums it up this way when describing its local models: “Powered by the LM Studio runtime, with MLX and llama.cpp under the hood.” Its documentation refers to .gguf and .safetensors files. When you download a model, you’ll see several versions with names like Q3_K_S or Q_8: they’re the same model, quantized, meaning compressed with more or less fidelity so it takes up less space.
Here are some of the models in the official catalog as of November 16, 2026, and where they run:
| Model | Maker | Where it runs | Review |
|---|---|---|---|
| Qwen3.8-27B | Alibaba | Locally (download) | Qwen |
| Mistral Small (24B) | Mistral AI | Locally | Mistral |
| Gemma 4 (5.1B to 31B) | Locally | Gemini, from Google | |
| gpt-oss (20B and 120B) | OpenAI | Locally | OpenAI |
| Muse Glimmer (30B) | Meta | Locally | Llama |
| DeepSeek V4 Flash (284B) | DeepSeek | Locally and in LM Studio’s cloud | DeepSeek |
| DeepSeek V4 Pro, Kimi K3, Muse Spark Open | DeepSeek, Moonshot AI, Meta | Only in LM Studio’s cloud (US) | DeepSeek · Kimi · Llama |
Beyond the catalog, you can search Hugging Face for any compatible model, even by pasting the full URL into the search bar. Which model fits on your machine depends on its size and the quantization you pick; I don’t publish a RAM-per-model table because I haven’t measured it.
Local server, API, CLI, SDKs and MCP
This is what turns LM Studio into more than a chat app:
- OpenAI-compatible server. It exposes
/v1/models,/v1/responses,/v1/chat/completions,/v1/embeddings,/v1/completionsand/v1/decisions, on port 1234 in the documentation’s example. To reuse code written for OpenAI, the docs say it’s just a matter of “switching up the “base URL”.” Codex works because LM Studio implements/v1/responses. - Anthropic-compatible too. It has a
/v1/messagesendpoint, and on January 30, 2026 it published a guide to using its models in Claude Code: “Run Claude Code with any local model using LM Studio’s Anthropic-compatible API.” - The
lmsCLI. MIT-licensed and bundled with the app (“lms ships with LM Studio”). You use it to load models, start the server or see what’s in memory from the terminal. - llmster. The no-UI mode: “LM Studio’s headless daemon for servers, cloud instances, and CI.” It installs with a single command on Mac, Linux or Windows.
- SDKs.
@lmstudio/sdkfor TypeScript (npm) andlmstudiofor Python (PyPI), both MIT-licensed. - MCP. Since 0.3.17 it acts as a host for MCP servers, both local and remote, with an
mcp.jsonthat follows Cursor’s notation. Its own warning: “Never install MCPs from untrusted sources.” An MCP server can run code and read your files. - LM Link (preview). It connects several machines running LM Studio or llmster so you can use the models on a powerful machine from another one, with end-to-end encryption over Tailscale. Tools that point to your local server, such as Codex, Claude Code or OpenCode, can use those remote models. It’s free for the duration of the preview; after that, according to its website, there will be free and paid plans.

Privacy: what leaves your machine and what doesn’t
This is what matters most when a company is considering local AI. The privacy policy says: “If you download and run models locally, none of your messages, chat histories, and documents are ever transmitted from your system.” And it adds that “the application does not include telemetry or user-specific tracking.”
- What does leave: the app version, your operating system and your IP address when you check for updates or download models.
- If you use Bionic’s cloud or web search: your requests are processed “transiently” and aren’t kept once they’re done. The Bionic cloud is hosted in the US, with Zero Data Retention (ZDR).
- No telemetry, no data to request. The policy itself warns that, because it doesn’t identify users, it can’t handle requests for a copy of your data or to delete it.
- What I couldn’t verify: whether Element Labs signs a data processing agreement or standard contractual clauses, or whether there’s an EU region. If you’re going to process personal data in the Bionic cloud, get it in writing first.
◇ My advice for sensitive data. If the data can’t leave the company, work only with downloaded models and skip the cloud and web search. LM Link doesn’t upload your chats to LM Studio’s servers: only the list of devices, so it can connect them.
LM Studio vs LM Studio Bionic: not the same app
On July 16, 2026, Element Labs launched LM Studio Bionic, which it calls “the AI agent for open models.” The announcement is clear: “Bionic is a new, separate app from LM Studio. For advanced low-level configuration, you can continue to use LM Studio alongside Bionic.” Its current version is 1.2.0, released November 5, 2026.
| LM Studio | LM Studio Bionic | |
|---|---|---|
| What it is | An app for downloading models, chatting with them and serving them through an API | An agent that does the work for you: it codes, and creates and edits documents, spreadsheets and slide decks |
| Where the model runs | On your machine (or another one of yours, with LM Link) | On your machine, through LM Link or in LM Studio’s cloud (LM Studio Secure Cloud) |
| Extras | Server compatible with OpenAI and Anthropic, CLI, SDKs, MCP | Voice dictation with local transcription (Voxtral, by Mistral AI) and web search |
| Price | Free | Free plan; the cloud is paid |
| Version | 0.4.28 (Oct 31, 2026) | 1.2.0 (Nov 5, 2026) |
Bionic cloud pricing
Official prices from the pricing page, checked on November 16, 2026. They’re in US dollars (USD) per month:
| Plan | Price | What’s included |
|---|---|---|
| Free | $0 | The Bionic agent, local models with llama.cpp and MLX, offline voice transcription, LM Link for up to 5 devices and limited web search |
| Bionic+ | $20/mo | Everything in Free, plus US-hosted open models (Kimi K3, GLM 5.3, DeepSeek V4 Flash and more), discounted bulk tokens, web search and page extraction |
| Pro | $100/mo | Everything in Bionic+, with 5× the usage limits and early access to new features |
| Teams | No published price | An organization with pay-as-you-go pooled credits or seat-based plans |
| Enterprise | Contact sales | SSO, SOC 2 Type 2 and deployment on your own infrastructure |
◆ Bionic doesn’t cost $20. Bionic has a Free plan, and the $20/mo is only for using models in its cloud. Anything beyond your plan’s allowance is paid with credits. Prices are only published in dollars, and I couldn’t confirm how taxes are applied, so check the final amount in your account before you budget. As for platforms, the homepage links Bionic downloads for Apple Silicon Macs, Windows (x64 and ARM64) and Linux (x64 and ARM64), but not for Intel Macs.
LM Studio vs Ollama: which one to pick
This is the most-searched comparison (“lm studio vs ollama” gets 260 searches a month in Spain and 3,600 in the United States, according to DataForSEO’s Google Ads data from November 16, 2026), and most of what’s out there dates from 2025. Here’s what I can back up with official sources:
| LM Studio | Ollama | |
|---|---|---|
| Code | Proprietary app (CLI and SDKs under MIT) | MIT-licensed repository |
| Local use | Free, including at work | Free: “Run models locally, No subscription” |
| Optional cloud | Bionic+ $20/mo · Pro $100/mo | Pro $20/mo ($16.67/mo billed annually) · Max $100/mo · Team $500/mo (early access) |
| Engine | llama.cpp and MLX | Not verified for this review |
| API | Compatible with OpenAI and Anthropic | Not verified for this review |
My call as an operator, based on those facts:
- Pick LM Studio if the person using it isn’t technical and wants an app with a catalog and a chat, if you’re on an Apple Silicon Mac and want to take advantage of MLX, or if you’re connecting tools that speak the Anthropic API, like Claude Code.
- Pick Ollama if your company requires open, auditable code. LM Studio’s app isn’t.
- Which one is faster? I don’t know, and I won’t make it up: I haven’t run my own test, and the ones floating around are third-party and mostly from 2025. If speed matters to you, test both on your own machine with the same model in each.
LM Studio alternatives
Data from the official websites and repositories, checked on November 16, 2026.
| Alternative | License | Price | Note |
|---|---|---|---|
| Ollama | MIT code | Free locally; cloud from $20/mo | I haven’t verified the license of its desktop app |
| Jan (Menlo Research) | Apache 2.0 | “Free & Open source” | Desktop app; version 0.8.8, released October 28, 2026 |
| llama.cpp | MIT | Free | One of LM Studio’s two engines; for anyone who wants to work with the engine directly |
| Msty Studio | Not verified | Free “$0 Forever” · Aurum $149 per user per year · Aurum Lifetime $349 per user (one-time payment) | Teams and Enterprise on request (5-seat minimum, with SSO) |
| LocalAI | MIT | Free | OpenAI-compatible API, built for servers |
| GPT4All | MIT | Free | I don’t recommend it today: its latest release is 3.10.0, from February 2025, and the repository hasn’t been touched since May 2025 |
What if you don’t want to run anything locally? Then the alternative isn’t another app but an API: OpenRouter gives you lots of models with a single key, although your data leaves your machine.
Who it’s for and who it isn’t
I’d recommend it if your company can’t or won’t send data to a cloud API, if your team needs a visual tool to work with local AI, or if you want your coding agents to use a model of your own without paying per token.
I wouldn’t recommend it if you’re going to sell an AI service to third parties built on the app (the license doesn’t allow it), if your team works on Intel Macs, if you need auditable open-source code, or if you’re after top-tier reasoning and don’t have the hardware: a 16 GB laptop fits small models, not the most powerful ones in the catalog.
My verdict as an operator
LM Studio is the most direct way I know to get local AI without touching the terminal: you download a model, chat with it and, whenever you want, start a server your tools can connect to as if it were the OpenAI or Anthropic API. Being free at work since July 2025 removes one of the main objections to using it professionally.
What I won’t hide from you: it isn’t open source, its license only covers personal or internal use, Intel Macs have been left out, and the brand has gotten confusing with Bionic, which is a different app with a paid, US-hosted cloud. My recommendation: for sensitive data, LM Studio with downloaded models and no cloud; Bionic only if you want an agent and accept what its cloud entails.
I’ll set up local AI for you and train your team
I can set up local AI for you if you don’t want your data leaving the building, and train the team that will use it. The work is choosing the model that fits your hardware, getting the app or the server up and running, connecting it to your tools and teaching you to maintain it without depending on me.
- If you want an agent built that runs on a model on your own machine: AI agent implementation.
- If you want me to teach you and your team to work with local AI: AI Mentoring.
- If you’re not sure where to start: tell me about your situation and I’ll tell you whether LM Studio is a fit or whether something else makes more sense.
LM Studio FAQ
What kind of software is LM Studio?
It’s a desktop app from Element Labs for downloading open AI models and running them on your computer. You use it to chat with them without sending data anywhere, query documents offline and serve the models through a local API that other tools connect to.
How does LM Studio work?
In three steps, according to its documentation: in the Discover tab you search for and download a model, you load it into memory from the model loader, and you chat with it in the Chat tab. If you want to connect it to other tools, you start the local server (in the official example, http://localhost:1234/v1).
Is LM Studio fully free?
Yes. Since July 8, 2025, it’s free at home and at work, with no commercial license to request. The only paid part is the cloud for its other app, Bionic: Bionic+ costs $20/mo and Pro $100/mo. Keep in mind that free doesn’t mean open source.
How much does LM Studio cost?
LM Studio itself costs nothing. If you use the Bionic app’s cloud, the plans are Free ($0), Bionic+ ($20/mo) and Pro ($100/mo), in US dollars. Teams and Enterprise have no published price.
Can I use LM Studio at my company?
Yes, for internal use. The license covers “personal and / or internal business purposes.” What you can’t do is resell it or offer it as a service to third parties. And every model you download has its own license.
What is LM Studio vs Bionic?
Bionic is a new, separate app launched on July 16, 2026: an agent that codes and works with documents, spreadsheets and slide decks, using models locally or in LM Studio’s cloud. LM Studio is still the app for downloading models, chatting with them and serving them through an API. You can use both side by side.
Which is better for Mac, LM Studio or Ollama?
On an Apple Silicon Mac, LM Studio can run models with MLX, the engine built for these chips, as well as llama.cpp; I haven’t verified whether Ollama does the same. Ollama’s advantage is that its code is MIT-licensed. On speed, I don’t have a test of my own I can cite. One thing to know either way: LM Studio doesn’t support Intel Macs.
What models can I use in LM Studio?
The ones in its official catalog (for example, Qwen3.8-27B, Gemma 4, gpt-oss, Mistral Small or DeepSeek V4 Flash locally) and any compatible model you find on Hugging Face (the documentation refers to .gguf and .safetensors files). Some large models, such as Kimi K3 or DeepSeek V4 Pro, are only available in its cloud.
What is the best local AI model?
There’s no single best one: it depends on the task, the language and how much memory your machine has. My advice is to try two or three catalog models that fit in your RAM on a real use case of yours. Before you use it at work, check the model’s license: my Qwen review covers the license of each model in that family.
Does LM Studio work on an Intel Mac?
No. The official documentation puts it this way: “Intel-based Macs are currently not supported.” You need an Apple Silicon Mac (M1 or later) running macOS 14.0 or later.
How much RAM do I need for LM Studio?
The official documentation recommends 16 GB or more. On a Mac with 8 GB it works with small models and modest context sizes. On Windows it also recommends at least 4 GB of dedicated VRAM and requires a processor with AVX2.
Is LM Studio 100% private?
Locally, yes: with downloaded models, no messages or documents leave your machine, and the app has no telemetry. It only sends technical data (version, operating system and IP address) when it checks for updates or downloads models. If you use Bionic’s cloud or web search, your requests go to US-hosted servers with zero data retention.
Can you trust LM Studio?
Here’s what I can verify: its privacy policy says nothing leaves your machine when you use local models and that the app has no telemetry. What I can’t do is audit it, because the app isn’t open source. Be careful with the MCP servers you install; LM Studio itself warns: “Never install MCPs from untrusted sources.”
Is LM Studio open source?
No. The app is proprietary, and its Terms prohibit reverse engineering. The lms CLI, the TypeScript and Python SDKs and the MLX engine are open source, under the MIT license.
Can I use LM Studio with Claude Code or Codex?
Yes. Claude Code connects through its Anthropic-compatible API, and Codex through its OpenAI-compatible /v1/responses endpoint. With LM Link, those tools can also use models on another machine of yours.
What is better than LM Studio?
It depends on what you need. To stay local: Ollama if you want open-source code (MIT), Jan if you want an open-source desktop app (Apache 2.0) and llama.cpp if you’re technical and want to go straight to the engine. I don’t recommend GPT4All today because it hasn’t been updated since 2025.
Can you help me set up local AI at my company?
Yes. I can set up local AI for you if you don’t want your data leaving the building, and train the team that will use it. You can see how I approach AI agents or message me directly.
Official sources checked on November 16, 2026: LM Studio download · 0.4.28 release notes · Changelog · Terms of Use (Aug 23, 2026) · Privacy policy · “LM Studio is free for use at work” · Bionic announcement · Pricing · Enterprise · System requirements · Getting started · OpenAI compatibility · Developer docs · MCP · lms CLI · llmster · LM Link · Model catalog · Blog · lms on GitHub · mlx-engine on GitHub. Alternatives: Ollama pricing and repository · Jan and repository · llama.cpp · Msty pricing · LocalAI · GPT4All.