How to Run AI Locally on a Laptop With No Internet: LM Studio and Open-Weight Models
LM Studio lets you run AI models offline to process sensitive documents securely. Learn how to set it up and use it for PII detection and compliance.
Why Running AI Offline Is a Serious Option Now
Running AI locally on your laptop — no internet, no cloud, no data leaving your machine — used to require a PhD and a server rack. That’s no longer true.
Tools like LM Studio have made it genuinely accessible to run open-weight AI models entirely offline. If you work with sensitive documents, handle personally identifiable information (PII), operate under compliance regulations like HIPAA or GDPR, or simply don’t want your data touching a third-party server, local AI is worth your attention.
This guide walks through exactly how to do it: how to install and configure LM Studio, which open-weight models to use, and how to run practical tasks — including PII detection and document review — without an internet connection.
What LM Studio Actually Is
LM Studio is a desktop application that lets you download and run large language models (LLMs) directly on your local machine. It supports Windows, macOS (including Apple Silicon), and Linux. You get a clean interface for chatting with models, a built-in model browser, and a local server you can query with API calls — just like OpenAI’s API, but running on your own hardware.
The key difference from cloud AI services: your prompts, your documents, and the model’s responses never leave your computer. There’s no telemetry, no third-party processing, no data retention policy to read through.
LM Studio runs “open-weight” models — AI models where the weights (the trained parameters) are publicly available for download and local use. These include well-known model families like Meta’s Llama, Mistral, Phi, Gemma, and Qwen.
Hardware Requirements Before You Start
Local AI does have real hardware demands. You don’t need a GPU, but it helps significantly.
Minimum Setup (CPU Only)
- At least 8GB of RAM (16GB recommended)
- A modern CPU — Intel Core i7 / AMD Ryzen 7 or better
- 10–20GB of free storage per model
- Works, but will be slow — expect 1–5 tokens per second on small models
Recommended Setup
- 16–32GB RAM
- A dedicated GPU with at least 8GB VRAM (NVIDIA RTX 3060 or better, or Apple Silicon)
- Apple M-series chips (M1, M2, M3, M4) are exceptionally well-suited for this — they use unified memory, so a 16GB or 32GB MacBook Pro can run 7B–13B parameter models quickly
- 50–100GB of free SSD storage
What to Realistically Expect
On a modern MacBook Pro with an M2 chip and 16GB RAM, a 7B-parameter quantized model generates around 30–50 tokens per second. That’s fast enough for practical use. On a mid-range Windows laptop with no dedicated GPU, the same model might run at 5–10 tokens per second — slower, but still functional.
The trick is using quantized models — compressed versions of larger models that sacrifice a small amount of accuracy for dramatically reduced memory requirements. More on that below.
Installing LM Studio
The installation process is straightforward.
-
Download LM Studio from the official LM Studio website. Choose the version for your operating system (macOS, Windows, or Linux).
-
Run the installer. No configuration required during installation. LM Studio installs like any other desktop app.
-
Open LM Studio. On first launch, you’ll see the main interface with a search bar for models, a chat window, and settings.
-
Check your hardware detection. LM Studio will automatically detect your GPU (if available) and configure itself accordingly. On Apple Silicon, it will use Metal acceleration by default. On NVIDIA GPUs, it uses CUDA. You can check this in the settings panel.
That’s the full install process. You don’t need to configure Python environments, install CUDA drivers separately, or touch a terminal.
Downloading Open-Weight Models
LM Studio has a built-in model browser that connects to Hugging Face, the primary repository for open-weight AI models. Here’s how to get a model onto your machine.
Step 1: Search for a Model
In the “Discover” tab, search for models by name or task. You’ll see metadata including model size, quantization level, and compatibility with your hardware.
Step 2: Understand Quantization
Models are listed with labels like Q4_K_M, Q5_K_M, Q8_0. These indicate quantization levels:
- Q4 = 4-bit quantization. Smaller file, faster, slightly reduced quality.
- Q5/Q6 = Good balance of size and quality.
- Q8 = Near full-precision quality, but larger file size and slower.
For most laptops, start with Q4_K_M or Q5_K_M variants. They’re practical.
Step 3: Download the Model
Click the download button next to the model variant you want. LM Studio handles the download and stores the model locally. Once downloaded, the model is available indefinitely — even with no internet.
Recommended Models for Sensitive Document Work
| Task | Model | Size (Q4) |
|---|---|---|
| General document analysis | Llama 3.1 8B Instruct | ~5GB |
| PII detection / compliance | Mistral 7B Instruct | ~4.5GB |
| Complex reasoning | Phi-4 (14B) | ~9GB |
| Fast, lightweight tasks | Gemma 2 9B | ~6GB |
| Structured output / JSON | Qwen 2.5 7B Instruct | ~5GB |
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
For most sensitive document workflows, a 7B–8B parameter model with Q4 or Q5 quantization is the right starting point. You can always test larger models if your hardware supports them.
Running Your First Offline AI Session
Once you have a model downloaded, here’s how to run it.
Starting the Chat Interface
- Go to the “Chat” tab in LM Studio.
- Click the model selector at the top and choose the model you downloaded.
- LM Studio will load the model into memory. This takes 5–30 seconds depending on model size and your hardware.
- Type your prompt and press Enter. That’s it.
At this point, your laptop is processing the request entirely on-device. No network traffic. You can verify this by enabling airplane mode — everything still works.
Setting Up the Local Server
LM Studio also runs a local API server compatible with the OpenAI API format. This is useful if you want to query the model programmatically.
- Go to the “Local Server” tab.
- Click “Start Server.” The default port is
1234. - You can now send requests to
http://localhost:1234/v1/chat/completionsusing the same format as the OpenAI API.
This means any tool or script that works with OpenAI’s API can be pointed to your local LM Studio server instead — with zero code changes in most cases.
Using Local AI for PII Detection and Compliance
This is where offline AI becomes genuinely valuable in professional settings. If you work with legal documents, medical records, HR files, or financial statements, you probably can’t send that content to a cloud API without triggering compliance concerns.
Local AI sidesteps that entirely.
Building a PII Detection Workflow
Here’s a practical prompt approach for scanning documents for personally identifiable information:
You are a compliance assistant. Review the following document and identify any personally identifiable information (PII), including:
- Full names
- Email addresses
- Phone numbers
- Social Security Numbers or national ID numbers
- Dates of birth
- Physical addresses
- Financial account numbers
- Medical record numbers
For each item found, return:
1. The type of PII
2. The exact text
3. The character position or surrounding context
4. A suggested redaction action
Document:
[PASTE DOCUMENT TEXT HERE]With a model like Mistral 7B Instruct or Llama 3.1 8B, this prompt reliably extracts PII from unstructured text. The output can be structured as JSON if you append: “Return your response as a JSON array.”
Document Review Without Data Exposure
The same approach works for:
- Contract review: Flagging unusual clauses, missing terms, or non-standard language
- HIPAA compliance checks: Identifying protected health information (PHI) in notes or records
- GDPR data mapping: Spotting personal data categories across documents
- Internal audit preparation: Extracting key facts, dates, and figures from dense documents
The practical workflow is simple: paste document text into the chat (or use the local API to automate it), run your prompt, get structured output.
Limitations to Know
Open-weight models at the 7B–8B scale are capable, but not infallible. For high-stakes compliance tasks:
- Always have a human review the output.
- Test your prompts against known examples before deploying.
- Larger models (13B, 32B) are more accurate but require more hardware.
- Structured output (JSON mode) reduces hallucination in extraction tasks.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Tips for Getting Better Results Offline
A few things that meaningfully improve output quality when running local models:
Use system prompts. LM Studio lets you set a system prompt that runs before every conversation. Use this to define the model’s role, output format, and constraints.
Keep context windows in mind. Most 7B models handle 4,000–8,000 tokens of context. Long documents may need chunking. Split a 20-page contract into sections and process each separately.
Adjust temperature settings. For extraction tasks (PII detection, fact-finding), set temperature to 0 or 0.1. Lower temperature means more deterministic, consistent output. For summarization or drafting, 0.3–0.7 works better.
Enable GPU layers. In LM Studio’s model settings, you can configure how many model layers run on the GPU. Maximizing GPU layers speeds things up significantly. The app will tell you if you’re running out of VRAM.
Test multiple models. Different models have different strengths. Mistral 7B often outperforms Llama 7B on instruction-following tasks. Qwen 2.5 models are strong at structured JSON output. Try a few with your actual prompts to find the best fit.
Where MindStudio Fits Into a Local AI Workflow
Running AI locally with LM Studio solves the data privacy problem. But building a full workflow around it — intake forms, routing, output formatting, notifications — still takes engineering time if you’re working from scratch.
This is where MindStudio connects. MindStudio’s AI Media Workbench and workflow builder natively support local model providers, including LM Studio and Ollama. You can configure a MindStudio agent to call your local LM Studio server (via its OpenAI-compatible API at localhost:1234) as the underlying model, then build the surrounding workflow visually — no code required.
What that looks like in practice: you set up a MindStudio agent that accepts a document upload, sends the text to your local LM Studio instance for PII detection, formats the results, and routes them to the right place (a Notion database, a Slack channel, a Google Sheet — MindStudio has 1,000+ integrations). The sensitive content never leaves your network; only the structured, sanitized output goes to downstream tools.
For teams that need to process sensitive documents regularly, this pattern — local model for the sensitive reasoning step, MindStudio for the surrounding workflow automation — covers a lot of ground without requiring a full engineering build.
You can start building on MindStudio for free at mindstudio.ai.
Frequently Asked Questions
Can I run LM Studio completely offline with no internet connection?
Yes. Once you’ve downloaded LM Studio and the model files you need, the application runs entirely locally. You can enable airplane mode and everything continues to work. The only time internet is required is during the initial app download and model downloads.
What is the difference between LM Studio and Ollama?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Both let you run open-weight models locally. LM Studio has a full GUI application and is easier to get started with, especially for non-technical users. Ollama is more developer-focused and primarily command-line driven, but it’s lighter weight and easier to integrate into automated pipelines. Both expose an OpenAI-compatible API. For document review work without coding, LM Studio is the more accessible starting point. For developers building automated workflows, Ollama often pairs better with scripts and applications.
Is LM Studio safe to use with sensitive data?
For the purpose of keeping data off external servers, yes — your data stays local. The application itself doesn’t send prompts or outputs to any external service. That said, you should still review LM Studio’s privacy policy and your organization’s security requirements. Running on an air-gapped machine (no internet at all) gives the highest level of assurance.
Which open-weight models are best for document analysis and compliance tasks?
For most users, Llama 3.1 8B Instruct and Mistral 7B Instruct are the best starting points. They’re well-documented, widely tested, and perform reliably on instruction-following tasks like PII extraction and document review. If your hardware supports larger models, Llama 3.1 70B or Qwen 2.5 72B offer significantly better accuracy but require 40GB+ of RAM or high-VRAM GPUs.
Can I use local AI models for production compliance workflows?
You can, but with caveats. Local open-weight models are useful tools for drafting, flagging, and triage — but they’re not certified compliance tools. They can miss PII, misinterpret context, or hallucinate. For anything with legal or regulatory weight, local AI output should be treated as a first pass, reviewed by a qualified human. That said, even as a draft-and-flag tool, they can dramatically reduce manual review time.
How much storage do I need for local AI models?
A typical 7B parameter model with Q4 quantization takes roughly 4–5GB of disk space. A 13B model takes 8–10GB. If you plan to experiment with multiple models, budget 50–100GB of free SSD space. Model files are stored in a single local directory that LM Studio manages for you.
Key Takeaways
- LM Studio makes local AI accessible. You don’t need to write code or configure environments. Download the app, download a model, and start using it — all offline.
- Open-weight models at 7B–8B scale are practical on modern laptops. Apple Silicon MacBooks and mid-range Windows machines with dedicated GPUs can run them well.
- Local AI is a real option for sensitive documents. PII detection, contract review, and compliance screening are all viable use cases — as long as you include human review of outputs.
- Quantized models are the practical choice. Q4 or Q5 variants run faster and use less memory with minimal accuracy loss.
- LM Studio’s local API server lets you integrate local AI into broader workflows. Tools like MindStudio can connect to it, letting you build automated document processing pipelines without exposing sensitive data to cloud services.
- Start simple. Try one model, one task, one document. You don’t need to build a full pipeline on day one.
