LM Studio

一款可以在个人电脑上本地, 私密地运行open-source large language model的免费桌面 app程序

large language model

https://lmstudio.ai/

Visit Website

In-depth Report

LM Studio is a free desktop application developed by Element Labs Inc. that allows users to run open source large language models (LLMs) on their local computer with no cloud connection or subscription fees. With its zero-threshold graphical interface design and comprehensive functional integration, this product has become a benchmark product in the field of local large model deployment. LM Studio supports mainstream open source models such as Llama, Qwen, DeepSeek, and Gemma, and provides OpenAI-compatible API interfaces, built-in RAG document retrieval, and cross-platform GPU acceleration capabilities. It is recommended as a key cooperation tool by many hardware manufacturers (NVIDIA, AMD, Apple). The latest version is 0.4.12. The main version iteration focuses on enterprise-level functional expansion such as headless deployment, parallel reasoning, and remote connection (LM Link).
LM Studio is developed by Element Labs Inc., whose LinkedIn profile is linkedin.com/company/elementlabsinc. The core members of the development team include yagil, mattjcly, rugvedS07, neilmehta24 and others. According to official recruitment information (lmstudio.ai/careers), the company is currently in the expansion stage and is recruiting talents. LM Studio has established an active open source ecosystem on GitHub. The organization's homepage is github.com/lmstudio-ai, with approximately 8,957+ Stars and 11 public warehouses. Major projects include: lms (CLI tool, 4,652 Stars), lmstudio-js (TypeScript SDK, 1,599 Stars), mlx-engine (Apple MLX engine, 1,002 Stars), lmstudio-python (Python SDK, 794 Stars). It is worth noting that the LM Studio desktop application itself is closed source software, but some open source tool chains are built around it. LM Studio has received official recommendations and technical support from multiple chip manufacturers. The official NVIDIA blog published an in-depth technical article in May 2025, introducing how LM Studio leverages RTX GPU and CUDA 12.8 to achieve performance optimization. AMD’s official website also includes LM Studio as an important partner of the Ryzen AI ecosystem. For Apple Silicon, LM Studio implements unified memory optimization for M series chips through the MLX engine.
Local inference and model management: LM Studio is built based on the llama.cpp high-performance inference engine and supports running a variety of open source large models such as gpt-oss, Llama, Qwen3, Gemma3, DeepSeek, etc. on local hardware. The product has a built-in LM Studio Hub model market, and users can directly browse and download thousands of GGUF format models on Hugging Face. After the model is downloaded, it is completely available offline. All inference processes are completed on the user's own device, and the data never leaves the local area. GPU acceleration support: Cross-platform GPU optimization is the technical highlight of LM Studio. NVIDIA integrates CUDA 12.8 to achieve up to 35% throughput improvement through computational graph optimization and Flash Attention CUDA kernel. Apple Silicon implements Metal acceleration through the MLX engine, taking full advantage of the unified memory architecture. Windows platforms support DirectML and CUDA. It's worth noting that LM Studio currently does not support AMD ROCm graphics cards. API and developer functions: The product provides an OpenAI-compatible REST API interface (default localhost:1234/v1), and an Anthropic-compatible interface (v0.4.7+). Developers can integrate through JavaScript SDK (@lmstudio/sdk) and Python SDK (lmstudio). Version 0.4.0 adds the headless daemon llmster, which supports deployment in Linux servers, cloud instances and even CI environments without the need for a GUI. The LM Link feature introduced in version 0.4.5/0.4.6 supports end-to-end encrypted remote instance connections. RAG and document processing: Built-in retrieval-enhanced generation (RAG) capability supports loading of local PDF, text, code and other documents for context-based dialogue. Version 0.4.0 enhances the chat export function and supports PDF, Markdown and plain text formats. Context window usage visualization is a highly rated feature by users, showing context fill percentage in real time.
v0.4.x series (2026): 0.4.0 (January 28) releases major updates such as headless deployment, parallel reasoning, new REST API, and comprehensive UI refresh. 0.4.5/0.4.6 (February) introduces the LM Link remote connection function. 0.4.7 (March) enables the separation of reasoning_content and content by default. 0.4.8 (March) adds reasoning_effort parameter support. 0.4.12 (April 17) supports Qwen 3.6 and optimizes the performance of Qwen 3.5 on various compatible interfaces. v0.3.x series (2024-2025): 0.3.4 (October 2024) introduces Apple MLX engine support. 0.3.5 (October 2024) adds headless mode. 0.3.16 (May 2025) launches community presets, LLM text embedding, KV Cache uninstallation and other functions.
Interface design: Users generally speak highly of LM Studio’s graphical interface. Product Hunt users say "the interface is beautiful and super easy to set up and get started." Reddit users stated that "the interface makes using local models much simpler than command-line tools, making it suitable for experimenting with different language models without consuming subscription points." The Elephas Blog review gave it a 4.3/5 score, and the ToolChase review gave it a 4.5/5 score. Contextual visualization: Several reviews singled out the contextual window usage display as a "game-changing" feature. Compared to ChatGPT, which does not issue a warning when the context is full, LM Studio clearly displays the context fill percentage, prompting the model to resolve core issues as quickly as possible before it is forgotten. Electron architecture issues: Feedback from experienced users pointed out that LM Studio is built on Electron and is a large resource consumer. Since running large models locally is extremely memory intensive, the additional 1-2GB of memory occupied by Electron is not friendly enough for devices such as the 8GB MacBook Air.
LM Studio adopts the business model of "free for individuals + customized for enterprises". According to the official statement, personal use is completely free, and enterprise use requires contacting the sales team to obtain a quote. This model has triggered some discussions in the user community, and some users have expressed dissatisfaction with the "no price indication" sales strategy. The free version includes all core features: model download and local chat, local API server, RAG document processing, Python and TypeScript SDKs. The paid enterprise version is geared toward team organizations and offers shared team inference endpoints, enterprise support, volume licensing, and dedicated account management.
Positive comments: Users’ recognition of LM Studio mainly focuses on the following aspects: fully local operation guarantees data privacy, zero subscription cost reduces the threshold for use, GPU acceleration improves inference efficiency, the graphical interface reduces the difficulty of use, and the OpenAI compatible API facilitates integration. An in-depth review article from Alex Kraieski pointed out that the core motivation for using LM Studio is data privacy - to avoid personal privacy or employer intellectual property leakage into the cloud service provider's training data or advertising portraits. Negative feedback: High hardware requirements - 16GB RAM minimum, 32GB recommended to run larger models, not friendly to users on a budget. MacRumors forum users reported poor experience on base Mac configurations, with speeds of only 5-8 tokens/s when running 70B models. Trust issues in closed source software - the source code cannot be viewed to verify the privacy statement, and users can only "trust" the product without collecting data. Almost all proprietary software collects telemetry data, and users who are extremely privacy-conscious may need to disconnect from the Internet. Other limitations include: no support for AMD graphics cards, lack of voice input capabilities, large model file sizes (3-40GB each), and possible server disconnection during long text generation.
Comparison of major competing products: LM Studio, Ollama, and Jan.ai form a three-legged competition. Ollama is the main competing product. It adopts a CLI-first design concept, its code is completely open source, and it is extremely active in the developer community (162k+ GitHub Stars). Ollama is suitable for developers, automated scripts, and multi-model concurrency scenarios. Jan.ai focuses on open source and privacy protection, adopts the AGPL protocol, provides a ChatGPT-like interface and a unique plug-in system, and is suitable for those who put privacy first. GPT4All is optimized for low-configuration hardware and provides a CPU-optimized version, suitable for non-technical users. vLLM has leading performance on NVIDIA GPUs and is suitable for enterprise-level high-concurrency services, but macOS support is limited. Selection recommendations: Non-technical users prefer LM Studio (best GUI experience); developers prefer Ollama (CLI priority, maximum control); privacy-sensitive users prefer Jan (completely open source, zero telemetry); Apple Silicon users who pursue performance can choose vMLX/oMLX.
Data privacy risks: Although LM Studio emphasizes that it runs completely locally and does not collect data, its desktop application is closed source software, and users cannot audit the code to verify the privacy statement. This means users must "trust" the provider not to collect telemetry data. For enterprise users who handle highly sensitive information, it is recommended to use caution after evaluating data governance requirements. Security Vulnerability: Users on the GitHub Bug Tracker reported that LM Studio executed MCP server commands before validating the MCP server handshake, which is a potential security vulnerability. Version 0.4.10 has added MCP server OAuth authentication support. Market positioning challenge: With the increasingly improved functions of competing products such as Ollama and Jan, and the rise of tools such as vMLX and oMLX specifically optimized for Apple Silicon, LM Studio's lead in "ease of use" is shrinking. At the same time, the closed-source strategy limits its appeal to privacy-sensitive users.
Suitable for: Privacy-conscious professionals who need to process sensitive data locally; users who pursue offline workflows and do not rely on cloud connections; non-technical users who want to use local LLM without command line operations; AI learners and researchers who need to quickly test and compare different open source models on the desktop. Not suitable for: Users with low hardware configuration (less than 8GB RAM) (GPT4All is recommended); Enterprises that require enterprise-level RAG functionality (AnythingLLM is recommended); Developers who prefer command line operations and pursue maximum control (Ollama is recommended); AMD graphics card users (LM Studio does not support AMD ROCm). Alternatives: Jan or privateGPT for privacy priority; vMLX/oMLX for Apple Silicon users pursuing performance; PremAI for enterprise compliance needs; Open WebUI + Ollama for multi-user scenarios.
LM Studio is a benchmark product in the field of local large model deployment. Its core value lies in encapsulating complex LLM running technology into a "zero-threshold" desktop application. With its exquisite graphical interface, rich model support, and complete API ecosystem, LM Studio has become the first choice tool for non-technical users to get started with local AI. However, limitations such as closed-source strategy, high hardware requirements, and lack of support for AMD graphics cards also restrict its user coverage. The release of version 0.4.0 marks the product's transformation from "desktop application" to "local AI development platform". Enterprise-level functions such as headless deployment and LM Link remote connection further expand the usage scenarios. For users who have privacy needs or wish to control their AI infrastructure, LM Studio is a solution worth considering, but it is recommended to evaluate hardware configuration and data security requirements before use.

User Reviews

CLbau

—

Bionic 出��以后我才开始认真考虑本地跑 Agent 这件事。不用绑云端 API，数据不出本机，还能在 Work Project 里处理 PDF 和表��，对我这种 freelancer 来说太合适了。

GeorgeBarnes_77

—

LM Studio 的 MLX 后端在 M4 Mac 上跑 Gemma 4 确实比 Ollama 快不少，吞吐能到 230+ tok/s。但内存占用也大，开一天风扇呼呼转，唉有利有弊吧。

Jordan_Ross168286

—

用了大半年了，从最开始的版本一路升上来。0.4 以后加了 llmster 命令行模式，终于可以在服务器上无头跑了，之前只能当桌面玩具被人诟病。虽然还是 preview 阶段，但方向对了。

DMurphyK

—

Bionic 那个 voice keyboard 挺惊艳的，用 Mistral Voxtral 离线语音转录，系统级部署，哪里打字切哪里。本地转录不上云，隐私方面确实比用 Siri 放心多了。

Adam.Flores

—

其实很多开发者拿 LM Studio 当模型测试台，找到合适的量化版本再丢去 Ollama 上��。两个工具搭配用挺常见的，不是非此即彼的关系。

cHAINlINK

—

LM Studio 界面是是真的漂亮，比 Jan 和 GPT4All 都精��。但闭源这件事始终让我有点膈应，哪天团队拿 VC 钱开始收割用户就难受了。

AHicks1689

—

下载模型实在太慢了，HuggingFace 的源在国外，我开梯子都经常断。好几次下到一半失败又要重来，几十个 G 的文件谁受得了啊。

LincolnGonzalez

—

刚入坑本地 AI 的时候折腾 Ollama 搞了一下午没跑通，朋友推荐 LM Studio，下载安装到跑起来十分钟不到。适合我这种只想用 AI 不想配环境的人。

TaylorGomez

—

用 LM Studio 的 API 服务器接 Cline 和 Continue 非常方便，改一下 base_url 就完事了。本地跑 DeepSeek 写代码，不用看 OpenAI 的脸色。

Gloria_Wilson_X

—

LM Link 出了以后我直接把家里 4090 的台式机当推理服务器，出门用 MacBook Air 远程连回来跑大模型。延迟可以接受，相当于随身带了个 4090。

tOKENmASTERpRICE

—

说实话 7B 以下的模型在 LM Studio 上跑起来体验还行，但一上 30B 以上的大模型，8G 显存直接爆炸。软件提示倒是做得好，下载前就告诉你 VRAM 够不够，避免折腾。

袁莉_1

—

我拿 LM Studio 做了个本地客服知识库，把公司几百页产品文档喂进去做 RAG。离线可用，数据不用上云，合规那边直接过了。老板说省了每年好几万的 API 费。

曹明静

—

对比测试过好多次了，在 RTX 4070 上同模型同参数，LM Studio 比 Ollama 慢大概 10%-15%。不过 GUI 确实好，每次调试 prompt 不用对着黑框输命令，滑块一拉就完事。

尹超心

—

用 LM Studio 的侧边栏调参数真的很直观，温度、Top-P、重复惩罚这些滑块一拉就看到效果变化。对于做 prompt engineering 的人来说省太多事了。

Andrew_Taylor520

—

免费给个人和商业用，这点确实良心。同类工具多少都有付费墙，LM Studio 完全免费还不限制功能，开发者团队靠什么活的我也挺好奇。

JMoralesQ0

—

LM Link 远程连 Mac 跑 iPhone 上用的体验还可以，就是切后台久一点会断连重连，有点烦。不过既然是端到端加密的，隐私方面确实靠谱。

VJames_Pro

—

中文用户劝退的一点是汉化还是 beta，很多界面还是英文。偏偏模型下载页面和参数说明又有不少中文术语混着，刚入门的看着有点懵。

CGarciaQ46

—

MCP 客户端支持加进来之后，本地模型终于能调用外部工具了。虽然配置还是有点麻烦要手写 JSON，但至少比之前只能纯聊天强多了。

JakubHaram

—

LM Studio 让我第一次在飞机上也能用 AI 写东西。离线下载好模型，高空没网络一样跑。相比之下那些云端 AI 断网就废了，出差党狂喜。

wu1oi

—

同一台电脑并排比较两个模型的输出，这功能谁用谁知道。有时候 A 模型写代码好但逻辑不严谨，B 模型严谨但啰嗦，摆在一起比就一目了然了。

OliviaLaurila

—

新版 0.4 那个 PDF 聊天有点意思，直接拖 PDF 进去就能问问题，不用搭 RAG 管道。虽然只是简单的检索不是向量索引，但应急够用了。

Debra.Cook_2021828

—

6G 显存的 2060 跑 7B 量化模型勉强能用，每秒十几个 token，慢是慢了点但至少跑起来了。LM Studio 的好处就是低配机也能有个入口体验一下本地 AI。

AmberChavez_Max6

—

我拿 LM Studio 处理公司机密合同摘要，数据不用出本机，法务那边很满意。之前用 ChatGPT 每次都心惊胆战的怕把敏感信息传出去。

MMoore_2022

—

本地跑的优势就是无限量随便用，不用心疼积分。我拿 LM Studio 跑了整晚批量翻译几百页技术文档，要不是电费就是零成本，云端 API 那得烧多少钱。

Keith.Lewis4

—

发现一个小坑，local server 同时只能服务一个模型。想切换得先卸载再加载别的，不像 Ollama 可以��时跑多个。做对比测试的时候有点不方便。

MBarnes520

—

装了LM Studio后彻底回不去用ChatGPT了，数据在自己电脑上跑的感觉太踏实了

Eric.Butler_2023045

—

界面真的好看，比Ollama那些命令行舒服太多

CherylPerry_Pro441

—

Qwen3.5跑起来速度还不错

尹婷阳

—

免费还不用申请API key，懒人福音

BDavis_7

—

用了一段时间，感觉最大的问题是太吃内存了，16G的Mac跑起来有点吃力

LM Studio

In-depth Report

User Reviews

Similar Products