By FreePortal Team · Reviewed by FreePortal Editorial Team · Last updated October 2026

How to Download and Install LM Studio Free Guide

Run powerful AI models like Llama 3.3, Qwen 3, and Gemma 4 directly on your PC — no internet or API keys required.

Is LM Studio Free?

Yes — for both personal and internal business use. LM Studio is free to download and use, and since 8 July 2025 it no longer requires a separate commercial licence. The EULA grants a licence for personal and internal business purposes at no cost. There is no subscription, no account requirement for local use, and no feature gate for the desktop app.

LM Studio is proprietary software, not open source — but it is free to use. What the terms prohibit is **reselling, redistributing, sublicensing, or building it into a product or hosted service you sell**. Running LM Studio on your own machine or a work laptop is explicitly allowed. Companies that need SSO, model gating, or private team collaboration can buy an **Enterprise** plan, but that is optional and aimed at larger organisations.

Two paid products sit alongside the free desktop app: **pay-as-you-go cloud inference** (for frontier open models too large to run locally) and **Bionic**, the agent app, which has a free tier plus **Bionic+ ($20/month)** and **Pro ($100/month)** tiers for cloud usage. Neither is required to run models locally.

System Requirements

Requirements below reflect the official LM Studio documentation (checked October 2026).

macOS: macOS 14 (Sonoma) or newer, on Apple Silicon only (M1–M4). Intel Macs are not supported. 16 GB+ RAM recommended; 8 GB Macs work with smaller models and modest context sizes.
Windows: Windows 10/11 on x64 with AVX2 support, or ARM64 (Snapdragon X Elite). 16 GB RAM recommended; 4 GB+ dedicated VRAM suggested for GPU offload.
Linux: Ubuntu 20.04 or newer, x64 (AVX2) or ARM64. Distributed as an AppImage; a `.deb` path exists as of 0.4.22. Versions newer than Ubuntu 22 are less thoroughly tested.
RAM: 8 GB minimum (16 GB+ recommended for 7B+ models)
GPU (optional): NVIDIA GPU with 4+ GB VRAM for faster inference; Apple Silicon Macs run well on unified memory
Storage: 10+ GB free space (models range from 2 GB to 40 GB+)
Note: Current release is 0.4.23/0.4.24 (August–September 2026). The 0.4.x line added parallel inference, a headless daemon, and a stateful REST API.

How to Install LM Studio

Go to the official site. Visit lmstudio.ai.
Download your version. Click “Download for Windows”, “macOS”, or “Linux” — the site auto-detects your OS. On Apple Silicon, there is also an official Homebrew cask: brew install --cask lm-studio.
Run the installer. Open the downloaded file and follow the prompts. No admin rights needed on most systems.
Launch LM Studio. Open the app from your Start menu or Applications folder.

Getting Started with Local Models

Download a model. In the app, open the Discover tab (⌘+2 / Ctrl+2) and search for a model like “Qwen 3 8B” or “Llama 3.3 8B”. You will see the same model at several quantization levels — choose Q4_K_M or higher.
Load the model. Go to “My Models”, click the settings icon on your downloaded model, and click “Load Model”. Set context length only as high as you need — the default is 8k tokens, and larger contexts consume significantly more memory.
Start chatting. Go to the “Chat” tab — you can now type prompts and get responses directly from your local AI.
Use the API server (optional). In the “Developer” tab, click “Start Server” to enable an OpenAI-compatible API at http://localhost:1234 (and an Anthropic-compatible /v1/messages endpoint) — great for connecting to Python scripts, Cursor, or VS Code.
Adjust settings. In Settings → Inference, tune GPU layers and context length to balance speed and memory. Start with GPU offload set to “max” if the model fits entirely in VRAM.

Model Recommendations by RAM

Choosing the right model size matters more than anything else in local AI. These are practical guidelines based on typical hardware, using current 2026 model families.

8 GB RAM — 3B–4B models: Qwen 3 4B, Gemma 4 2B, Phi-4 Mini. Fast and capable for general chat, but limited reasoning depth. Use Q4 quantization.
16 GB RAM — 7B–8B models: Llama 3.3 8B, Qwen 3 8B, Mistral Small 3.2. The sweet spot for quality vs. speed. Q4_K_M is the recommended default.
32 GB RAM — 13B–27B models: Qwen 3 14B, Gemma 4 27B, Mistral Nemo 12B. Noticeably better reasoning and instruction-following than 7B.
64 GB RAM (or 24 GB+ VRAM) — 30B–35B models: Qwen 3 32B, Qwen 3.8 27B, Mixtral 8x7B. Close to commercial model quality for most tasks.
128 GB RAM (or 2× 24 GB VRAM) — 70B models: Llama 3.3 70B, Qwen 3 72B. The largest open models that fit on consumer hardware. Slow without a GPU.

Start one size below what you think your hardware can handle — you can always download a larger model later once you know what works. LM Studio's own guidance is to choose **4-bit quantization or higher**; going below 4-bit to squeeze a larger model onto the same RAM usually produces worse results than running a smaller model properly.

Useful Tips

🧠 Start small. If you have 8 GB RAM, begin with 3B–4B models like Qwen 3 4B or Gemma 4 2B. For 16 GB RAM, try Qwen 3 8B or Llama 3.3 8B.
🚀 Enable GPU offload. In Settings → Inference, set GPU layers to “max” if the model fits entirely in VRAM. Partial offload is the usual cause of mysterious slowness — if even a few layers spill to CPU, throughput collapses.
💾 Download models once. Models are saved locally — you can load them anytime without re-downloading.
🔌 Connect to code editors. Use the local API server to integrate with VS Code, Cursor, or Python scripts — no API keys needed. See our Cursor guide if you use Cursor.
🔄 Stay at 4-bit or higher. Q4_K_M is a good balance; Q5_K_M gives better quality but uses more RAM; Q8_0 is closest to the original but requires significantly more memory. Avoid Q3 unless you have no other option.
🔒 Private and offline. Everything runs on your machine — no data leaves your PC, perfect for sensitive or confidential tasks.
⚙️ Understand context size. The default is 8k tokens as of 0.4.16. Larger context windows let the model remember more but consume significantly more RAM through the KV cache. Raise it only when you actually need long documents.
📊 Watch the load indicator. When you load a model, LM Studio shows how much of it fits on GPU vs. CPU. A model that fits fully on GPU responds much faster than one that spills to CPU.
🏢 Use it at work for free. The EULA allows internal business use at no cost. Only reselling or embedding it in a product requires an Enterprise arrangement.

LM Studio vs Ollama vs GPT4All

LM Studio: Polished GUI-first experience with model discovery, GPU/CPU offload controls, an OpenAI-compatible API server on port 1234, and an Anthropic-compatible endpoint. Adds parallel inference and a headless daemon in 0.4.x. Proprietary software, but free for personal and internal business use. Best for users who want a visual interface and are exploring local AI for the first time.
Ollama: Terminal-first with a simple CLI, excellent developer integration (Cursor, VS Code, Continue), and an OpenAI-compatible API on localhost:11434. MIT-licensed and open source, forkable and embeddable. Best for developers and anyone who wants to automate local AI from scripts or ship it inside a product. See our Ollama guide.
GPT4All: Free, open source, and lighter than both. Bundled with a local document chat feature (RAG) and model discovery. Emphasises CPU inference and is the simplest “install and chat” option. Best for users who want a basic local chat app without much configuration.

If you want a GUI and are exploring local AI, LM Studio is the best starting point — and it is now free at work too. If you want to build with local models from scripts or ship them inside a product, Ollama's MIT licence is the safer foundation. GPT4All is the lightest and simplest option for pure chat.

Troubleshooting Common Issues

❌ Model won't load — “out of memory” error. The model is too large for your RAM+VRAM. Try a smaller model (7B instead of 13B), a lower quantization (Q4 instead of Q6), or reduce GPU layers in Settings → Inference so more of the model stays on CPU (slower but works).
🐌 Responses are extremely slow. The model is running on CPU instead of GPU, or it is partially offloaded. Check Settings → Inference and enable GPU offload. If your GPU is enabled but still slow, reduce GPU layers — if the model exceeds VRAM, it spills back to CPU and runs slower than if it were all on CPU.
💾 Download fails or stalls. Model downloads go through Hugging Face's CDN, which occasionally throttles. Try a smaller quantization or a different model. If downloads consistently fail, check your network — corporate firewalls and VPNs sometimes block Hugging Face.
🔌 API server not responding. Confirm the server is running (check the Developer tab shows “Server is running”). Check your firewall isn't blocking port 1234. Test with curl http://localhost:1234/v1/models in a terminal — a response means the server is working.
🎨 Chat output is cut off mid-sentence. This usually means the model hit its context or output-token limit. Increase “Max Tokens” in the chat settings, or raise the context length when loading the model. If the context window is exhausted, start a new chat.
🖥️ GUI feels laggy. LM Studio's interface runs on top of the inference engine — if the machine is heavily loaded, the UI slows down too. Close other apps, or consider using the headless daemon (llmster) if you only need the API.
🍎 LM Studio won't install on an Intel Mac. Intel Macs are not supported. LM Studio requires macOS 14+ on Apple Silicon (M1–M4). There is no workaround — use a different local-inference tool on Intel hardware.
🧩 Context feels too short even at default. The default is 8k tokens. If you are working with long documents, raise the context length when loading the model — but expect memory use to rise sharply. For long-context tasks, a smaller model with a larger context often beats a larger model with a truncated context.

Frequently Asked Questions

Can I run LM Studio without a GPU?

Yes, but it will be slower. LM Studio falls back to CPU inference, which works but produces responses in tens of seconds rather than seconds. For occasional use, this is fine; for regular use, an NVIDIA GPU (8+ GB VRAM) or Apple Silicon Mac is much faster. LM Studio often outperforms Ollama on machines without dedicated GPUs because of its Vulkan offloading path.

What is the difference between LM Studio and Ollama?

LM Studio has a polished GUI and is easier for beginners to explore models visually. Ollama is terminal-first with excellent developer tooling and is fully open source under MIT. Both run the same underlying models and both expose an OpenAI-compatible API (LM Studio on port 1234, Ollama on 11434). LM Studio also adds an Anthropic-compatible /v1/messages endpoint and a native stateful REST API. See our Ollama guide for the developer-oriented alternative.

Can I use LM Studio commercially?

Yes — free for internal business use. LM Studio dropped its separate commercial licence on 8 July 2025. The current EULA grants a licence for personal and internal business purposes at no cost — no form or purchase required. Running LM Studio on a work laptop is explicitly allowed. What is not allowed is reselling it, redistributing it, or building it into a product you sell as a service — that requires an Enterprise conversation.

How big are the model downloads?

It depends on model size and quantization. Small 3B–4B models at Q4 are around 2–3 GB. 7B–8B models at Q4_K_M are around 4–5 GB. 13B–14B models are 8–10 GB. 30B–35B models are 18–22 GB at Q4. 70B models at Q4 are 40 GB+. Plan storage and download time accordingly.

Does LM Studio send my data anywhere?

No, for local models. Prompts, responses, and inference happen on your machine. The only network traffic is when you download models from Hugging Face, which goes through LM Studio’s CDN. After download, no data leaves your PC. The optional Bionic agent app and Secure Cloud inference are separate products with their own billing and zero-data-retention policies.

Can I use local models with Cursor or VS Code?

Yes, via the API server. Start the server in the Developer tab (OpenAI-compatible at http://localhost:1234), then point Cursor or the Continue extension in VS Code at that endpoint. This gives you AI coding assistance without API keys or data leaving your machine.

What is LM Studio Bionic?

Bionic is a separate AI agent app launched on 16 July 2026. It is built for coding, document work, slides, and local voice transcription. The free tier includes the Bionic agent, local models, offline voice transcription, and LM Link for up to five devices. Paid tiers — Bionic+ ($20/month) and Pro ($100/month) — add cloud inference with zero data retention and higher usage limits. Bionic is not required for local chat; the classic LM Studio app remains free.

← Back to all resources