LM Studio
Version in force 0.4.26-42026-10-08Tree audited closed source: no tree to auditLicence Proprietary
LM Studio 0.4.26-4 runs models on your computer without a terminal: Windows, Linux and Apple Silicon Macs, not Intel. It claims no telemetry, but its code is closed and its terms forbid reverse engineering: nobody outside can check, us included. By its own account, its 2 calls home out of the box are both for updates (app and engines).
A desktop application with a full graphical interface: search for models, download them, chat, and start an OpenAI-compatible server without ever opening a terminal.
Steps to the first token
Vendor declared
Hosts it contacts out of the box
Vendor declared
Licence
Vendor declared
Videos: LM Studio 3
The best walk-throughs we found. Press play and it opens right here.
The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.
How we checked this
Here we have to be upfront: everything this sheet says about the network comes from what the vendor declares — its privacy policy, its terms and its documentation — and from reading the code of the parts it does publish as open source (the lms command line, the SDKs and its MLX engine), on the date shown. The desktop app itself is closed source and its terms forbid decompiling it, so whatever the app does on its own can only carry the label «vendor declared», never «verified by us» — however curious we are. We have not run it in front of a network analyser either. We would rather tell you that than let you assume we checked what nobody outside the company can check.
What leaves your machine
Before a local runtime gets anywhere near your documents, we ask it four questions: what it sends on its own, who else it calls, which address it listens on, and whether you can check any of that yourself. Think of it as a job interview.
Telemetry out of the box
It states that it has none — and it is the only one of the five where nobody outside can check. Vendor declared, not independently verified
This is the tension at the heart of this comparison. LM Studio says in its privacy policy that the application contains no telemetry. We are not saying they are lying; we are saying the source is closed and their own terms forbid reverse engineering, so nobody outside the company can verify it. Four of the five rows in our table say verified; this one says vendor’s word. The definition is also narrow: the same policy admits that every update check sends app version and build, operating system and IP address, with the declared purpose of understanding the spread of their user base, and that they receive model search queries.
«Because the application does not include telemetry or user-specific tracking, we are unable to fulfill data subject requests such as providing a copy of your data or deleting your information.»
How to switch it off: Not applicable, according to the vendor: there is no documented telemetry switch, because in their account there is no telemetry to switch off.
Source · consulted 2026-08-25
Other calls home
| What | Where to | When | Can you avoid it | Evidence |
|---|---|---|---|---|
| Model search and download | huggingface.co |
Whenever you use the Discover tab. Their policy adds that they receive anonymised search queries. | Do not use it | vendor declared |
| Download and auto-update of the inference engines | LM Studio infrastructure |
On its own: engine auto-update is on out of the box (since 0.3.9). Also whenever you check for or install engines. | Turn engine auto-update off in the app settings | vendor declared |
| LM Studio Hub — presets and publishing | lmstudio.ai |
Only while signed in. | Do not sign in | vendor declared |
| LM Studio Cloud and web search | LM Studio infrastructure and its providers |
Only if you buy the cloud features and use them; web search requires signing in. | Do not buy them | vendor declared |
| LM Link — sharing models between your own machines | end-to-end encrypted network over Tailscale, discovery through the Hub |
Only if you turn it on. End-to-end encrypted over Tailscale, discovery through the Hub. | Yes, a setting turns it off | vendor declared |
«Device and app information (app version and build, operating system) and IP address (included automatically by our CDN provider). […] to give us an understanding of the software and hardware spread of our user-base.»
Source · consulted 2026-08-25
Which address it listens on
| Setting | Out of the box |
|---|---|
| Listens on | 127.0.0.1:1234 |
| Authentication out of the box | none |
| CORS out of the box | Disabled — of the five, only it and vLLM (since 0.32.0) ship with CORS closed |
Here is the twist that makes this table worth building: the one you cannot audit is the best configured out of the box. It listens on localhost only, it ships with CORS disabled (of the other 4, only vLLM does too, since 0.32.0, and vLLM listens on every interface), and it is the only one with an API-key system with scoped permissions. Its own network-serving page warns that any bind other than 127.0.0.1 exposes the server beyond localhost and recommends turning authentication on. Vendor declared, not independently verified
Source · consulted 2026-08-25
Can this be audited?
No. The application is closed source and its terms — the version dated 23 August 2026 — describe the code as a trade secret and expressly forbid reverse engineering, decompiling or disassembling it. Some pieces around it are open and MIT: the `lms` command line, the TypeScript and Python SDKs, the MLX engine and venvstacks. The program itself is not. That is why every telemetry claim on this sheet is the vendor’s word rather than a fact — not because we think it is false, but because there is no way to check. Verified by us in the primary source
«reverse engineer, decompile, disassemble, or otherwise attempt to derive the source code for the Software»
Source · consulted 2026-08-25
Privacy policy
Element Labs, Inc. (Delaware), policy dated June 2026. It states that if you download and run models locally, none of your messages, chat histories or documents ever leave your system. The Hub has its own, separate policy, which does collect email, profile name, IP address, browser type and actions taken. vendor declared Source
What it runs, and where
Platforms, accelerators and model formats: the three walls you can hit on that first, eager evening. Better to find them here than at midnight with a model half downloaded.
Where it runs
| Platform | Support | Evidence |
|---|---|---|
| Windows | Yes: x64 and ARM64 (Snapdragon X Elite). On x64 it requires the AVX2 instruction set. | verified by us |
| macOS Apple Silicon | Yes: macOS 14.0 or newer; they recommend 16 GB of memory or more. | vendor declared |
| macOS Intel | No. The documentation says Intel Macs are not supported, and their own download service returns HTTP 400 for darwin/x64 while darwin/arm64 serves the .dmg. | verified by us |
| Linux | Yes, as an AppImage or a .deb package, x64 and ARM64, Ubuntu 20.04 or newer. No .rpm. | verified by us |
What it accelerates
| Backend | Status | Evidence |
|---|---|---|
| CUDA | Yes, via a llama.cpp engine built with CUDA. | vendor declared |
| ROCm | Yes, via a llama.cpp engine with ROCm. | vendor declared |
| Metal | Yes: Metal through llama.cpp, plus its own MLX engine on Apple Silicon — and that engine is public and MIT. | vendor declared |
| Vulkan | Yes, via a llama.cpp engine with Vulkan. | vendor declared |
| CPU | Yes. On x64 it requires AVX2. | vendor declared |
| NPU | We could not verify this. No primary source mentions NPU support one way or the other, so we make no claim. | not verified |
Model formats it runs
GGUF and MLX. Its catalogue format abstracts both away. Whether it takes unconverted safetensors is not something we could verify. vendor declared
Source · consulted 2026-08-25
Installing it, and living with it
The everyday part: how many steps to the first token, what it leaves on your disk, under which licence, and at what pace it changes under your feet once it is installed.
Steps to the first token
3 steps. Three steps inside the application: search in the Discover tab, download, load and chat. No terminal at any point. vendor declared
Step by step: install it, start the server, check it answers, and lock down what can be locked down
If something fails: 7 errors, the exact message each one shows and the fix
What it writes to your disk
| What | Where | Evidence |
|---|---|---|
| Models | ~/.lmstudio/models/<publisher>/<model>/ — the directory is configurable |
vendor declared |
| Chat history, in plain JSON files | ~/.lmstudio/conversations/ |
vendor declared |
| Local server logs and Hub presets | ~/.lmstudio/server-logs/ · ~/.lmstudio/hub/ |
vendor declared |
Licence
Proprietary (EULA). Free for personal use and for internal business use since 8 July 2025 — but you may not resell it as a service: the terms forbid service-bureau use, application service providers and software-as-a-service. vendor declared
Source · consulted 2026-08-25
Source · consulted 2026-09-27
Version and release pace
Two to three releases a month: ten versions in four months, from 0.4.12 in April 2026 to 0.4.21 in August. Its sibling Bionic moves far faster — ten versions in 33 days. vendor declared
Source · consulted 2026-08-25
What the version in force brings: 0.4.26 (build 4, 8 October 2026) comes with, among other things, bug fixes and security hardening it does not detail, the llama.cpp 2.50.0 engine, peaceful coexistence with LM Studio Bionic (both open at once) and two new endpoints, /v1/decisions and /v1/systemone. And the bit we care most about here: none of its notes touches the system requirements, the port, the update check or telemetry. vendor declared
Source · consulted 2026-10-09
What to know before you install it
- It is the only one of the five whose network behaviour cannot be checked: closed source, and reverse engineering forbidden by contract. Its promise of no telemetry may well be true, but it is the vendor’s word, not a verifiable fact.
- Its own policy admits that every update check sends app version and build, operating system and IP address, and that they receive model search queries. We found no documented switch to stop that check.
- It is, on the other hand, the best locked down out of the box in everything that is configurable: localhost only, CORS off, and API keys with permissions.
- Careful with the marketing: the free plan card on their pricing page says «No data ever leaves your device» and, in the same card, lists a web search tool with the footnote «when logged in» — a tool that by definition sends the query out. Not a lie; imprecise.
- The product has forked: since July 2026 LM Studio Bionic is a separate application, Windows x64 and macOS ARM64 only. When somebody says «LM Studio» today, they may mean two different programs. This sheet is about LM Studio, not Bionic — and everything it says about the network was checked on 0.4.21-2, the version of 25 August 2026, not on the newer one shown above as the version in force.