LM Studio: a practical guide to installing and using local AI
LM Studio: a practical guide to installing and using local AI
LM Studio lets you download and run AI models on your computer through a chat interface and a local API. Once the model and required runtime components are downloaded, you can chat and work with documents without an Internet connection. Response speed, answer quality and usable document size depend on your hardware and the model you select.
“Free, offline and private AI” is an appealing proposition. Before relying on it, answer three practical questions: can your computer load the model, does it answer your document questions correctly, and does the entire workflow stay local? This guide shows how to check without buying a graphics card first.
Checked against official documentation on October 17, 2026. Features and prices may change. Examples are illustrative, not client results.
What is LM Studio, and what do you need to download?
LM Studio is the application that loads and runs a large language model, or LLM. The model is a separate set of files containing the AI’s learned parameters. Installing the application alone does not give you an offline assistant: you also need a model.
Available model families include Llama, Qwen and Mistral. Choose according to compatibility, licensing and available resources rather than brand recognition alone. LM Studio supports GGUF models through llama.cpp and MLX models on Apple Silicon Macs. You do not need to master these formats to start; the distinction explains why not every download fits every computer.
Check which product you are downloading. The current website also promotes LM Studio Bionic, an agent offering with local features and optional cloud services. This guide covers LM Studio’s local chat and document workflows. A cloud model or web search changes the offline and privacy assumptions.
A local model does not automatically become an agent that can work in your business applications. APIs and connected tools can enable that, but access and actions need configuring and testing. Our Hermes Agent business guide explains the difference between producing an answer and carrying out an assignment.
What hardware does LM Studio require?
Memory is the first constraint. The computer needs room for the loaded model as well as the conversation and document context. A model file can fit on your drive and still be too large to use comfortably.
- Mac: the documentation requires Apple Silicon and macOS 14 or later. It recommends at least 16 GB of memory. An 8 GB Mac may work with smaller models and modest context sizes. Intel Macs are not supported in the current requirements.
- Windows: supported x64 or ARM systems; x64 processors need AVX2 support. The vendor recommends at least 16 GB of RAM and a GPU with at least 4 GB of dedicated video memory, or VRAM.
- Linux: an AppImage is provided for x64 and ARM64. The official page lists Ubuntu 20.04 or newer, while warning that versions newer than 22 are less well tested. Check your environment rather than treating that statement as a guarantee for every distribution.
These official system requirements do not promise fast inference with a large model. RAM, VRAM, available acceleration, model size and context length all matter. On a work computer, browsers and business applications also consume memory; installed capacity is not the same as available capacity.
For a first trial, start small. Select a compatible model that leaves memory headroom. Quantization reduces memory requirements by representing model parameters at lower precision; its effect on answer quality depends on the model and task. Test the answers you need before deciding whether better hardware is justified.
How do you install LM Studio and run your first model?
- Download the application from the official download page. Choose LM Studio for your operating system, distinguishing it from Bionic if multiple products are offered. Install it and check the system requirements.
- Open Discover and select a model. Check its publisher, license, language capabilities and size. Downloadable weights do not mean unrestricted business use.
- Download the model while connected to the Internet. Allow the indicated disk space, plus room for any required runtime components. Trying several models means storing several sets of files.
- Select and load the model in Chat. Loading allocates memory. If it fails or makes the computer unresponsive, try a smaller model or a shorter context before changing unrelated settings.
- Ask a simple question, then test a realistic task. For example, rewrite a fictional customer email while preserving every date and amount. Check those details yourself.
- Verify the offline workflow. After all downloads are complete and the model is loaded, disconnect the network and repeat the task without external tools. This tests the setup you actually intend to use.
This sequence follows the official getting-started guide. Interface labels can change; the useful order is still application, model download, model loading and a checked response.
Does LM Studio really work offline?
Yes, for local functions prepared in advance. According to the offline-operation documentation, chatting with downloaded models, processing documents locally and running a local inference server do not require Internet connectivity.
Searching for models, downloading weights, fetching runtime components and checking for updates require network access. Catalog details and download options also make external requests. An offline assistant does not automatically know current news or newly published web content.
Adding a connected tool, remote MCP server or cloud model changes the data path. MCP is a way to connect tools; it does not mean that those tools run locally. Check where each request goes before describing the whole workflow as offline.
How do you chat with PDFs, Word documents and text files?
LM Studio accepts PDF, DOCX and TXT files in chat sessions. If a document fits within the model’s context—its working memory—the full text can be added to the conversation. For longer documents, the application can retrieve relevant passages and provide them to the model. This is retrieval-augmented generation, or RAG.
Retrieval can miss the right passage. A fluent answer is not evidence that every page was read or that the conclusion is correct. The document-chat documentation recommends asking specific questions with terms likely to appear in the source.
A useful test: attach a fictional procedure and ask, “What deadline applies to a request received on Friday? Quote the supporting passage, and say if the document does not provide an answer.” Compare the deadline and quotation with the original. Also ask a question the document cannot answer to see whether the model invents one.
A scanned PDF without extractable text may need optical character recognition first. With complex tables, check row alignment and units as well as the final wording. Start these trials with non-sensitive files, not a live customer dossier.
Which tasks are a good fit for local AI?
- Writing and rewriting: draft an email, outline or alternative wording, then review facts and tone.
- Exploring a small document collection: find instructions or compare passages, with a return to the original sources.
- Prototyping an AI feature: call a local model from a script or application through the available APIs. API compatibility does not mean complete equivalence with a cloud service.
- Working without connectivity: use models and files already on the computer, after testing the entire workflow before travelling.
Dropping files into a chat does not establish an up-to-date company knowledge base, team permissions or reliable answers across your entire history. When several people need shared, approved references, knowledge structure becomes a separate requirement. Our guide to knowledge graphs and business memory explains that larger design problem.
Privacy: what does local processing protect?
The vendor states that messages and documents processed by the documented local features stay on the device. In that scenario, their contents are not sent to a remote model API. This is a specific data-flow benefit, not a security certification for your entire computer.
Saved conversations, source files, computer backups, synced folders and other users’ access still need managing. A compromised laptop can expose information without making a cloud-model request. If you enable a server on the local network, every machine allowed to reach it becomes part of the security boundary.
The current pricing page also describes cloud services with zero data retention. No retention is not the same as no transmission. Do not confuse that offering with offline inference. Before handling regulated or confidential files, review the actual workflow, applicable terms and file controls. Running locally does not, by itself, establish GDPR compliance.
Is LM Studio free, and what does local AI actually cost?
The official pricing page lists a free offering that includes local model execution. At the time of checking, it also lists Bionic+ at USD 20 per month and Pro at USD 100 per month, with cloud services and different usage limits. Those subscriptions are not presented as a requirement for purely local chat. Check the plan that applies to the product you use.
When inference runs entirely on your computer, there is no remote API token bill for that inference. Hardware, electricity, storage, setup and review time remain real costs. A model’s license is a separate check from the price of the application.
An existing computer may be enough for one person rewriting a few texts. Multiple colleagues, long documents or strict response-time expectations can change the hardware and operating budget. Compare the total cost of completing the same work, rather than a cloud subscription against a free download.
LM Studio or a cloud assistant: how should you choose?
Start with a local trial if offline operation or limiting data transmission is a priority, and you can manage model selection and upkeep. A cloud service may be simpler for multi-user access or workloads beyond your machine’s capacity, provided its processing terms and costs are acceptable.
Use the same examples for both options: factual accuracy, required languages, long documents, waiting time and human review effort. A small local model can be useful for one bounded task and inadequate for another. If waiting time or errors make your examples too costly to review, choose a different setup before deploying.
A checklist before using LM Studio at work
- The operating system is supported and available memory is sufficient for the selected model.
- The model license allows the intended use.
- The workflow works with the network disconnected and no required external service.
- Answers have been checked against non-sensitive documents, including questions with no available answer.
- Conversation storage, backups and computer access are controlled.
- A named owner knows which outputs require review and which tasks must not be delegated.
- The budget includes hardware, maintenance and review time.
For Say Digital, the useful first deliverable is a documented trial: one task, approved test files, acceptance criteria and a go/no-go decision. If it succeeds, integration follows with regression tests, version control and reversible deployment. Without those checks, changing the model can change the answers the team relies on without anyone catching the regression.
LM Studio FAQ
Do you need a dedicated graphics card?
Not for every setup: acceleration and hardware resources differ by platform. On Windows, the vendor recommends at least 4 GB of dedicated VRAM. That does not mean any model will run smoothly within that amount. Check the official requirements and begin with a small model.
Can LM Studio run with 8 GB of RAM?
The documentation allows for smaller models and modest context sizes on 8 GB Macs, while recommending 16 GB or more. Do not generalize that statement to all computers or long document workflows.
Does attaching documents train the model?
Attaching a file supplies context or retrieved passages. It does not retrain the model’s learned parameters, nor does it guarantee that all later conversations will remember the document.
Can LM Studio replace ChatGPT?
It can cover some local chat, writing and document tasks. It is not a blanket replacement in features or answer quality. Your selected model, hardware and enabled services determine what is actually available.
Test local AI against a real business task
Want to know whether LM Studio fits your documents or an internal tool? Say Digital helps scope a trial using approved data, check the answers and compare costs before integration. You keep the deployment decision.
Official sources
Accessed October 17, 2026. Capabilities and recommendations above are vendor-documented, not the results of an independent audit.
