Local AI
Your computer.
Your AI.
Find models you can run on your own computer. Start with your memory, then choose what you want to do.
New to local AI? Get set upFind your next assistant for writing, coding, reasoning, or images.
Choose your installed memory to see a practical starting shortlist.
Where do I find my memory?
Windows: Settings → System → About → Installed RAM.
Mac: Apple menu → About This Mac → Memory.
NVIDIA GPU memory is separate from system RAM; don’t add the two together.
Starting estimates for compressed models with short conversations, allowing room for your operating system. Longer chats need more memory. GPU memory and your processor affect speed; RAM alone doesn’t guarantee performance.
Find your local model
16 models · Estimated to fit 16 GB RAM
Qwen3.5 · 2B
A small multimodal starting point for short drafts and simple image descriptions.
Qwen3.5 · 4B
Try everyday writing, image questions, and focused reasoning without a large model download.
Qwen3.5 · 9B
A middle-size everyday assistant for text, images, and short coding questions.
Gemma 4 · E2B
A compact Gemma4 entry point for writing and image questions on a personal laptop.
Gemma 4 · E4B
An everyday local multimodal option with configurable thinking on a laptop with some memory headroom.
LFM 2.5 · 8B
An on-device assistant for short instructions, summaries, and structured tasks.
Granite 4.2 · 8B
A current IBM text model for summarizing, extracting information, and explaining code.
Ministral 3 · 3B
A small multilingual model for text and simple image questions on an everyday laptop.
Ministral 3 · 8B
A balanced Mistral option for everyday writing, multilingual questions, and image understanding.
DeepSeek R1 0528 · 8B
The updated May 2025 Qwen3-based distill for trying local math, logic, and code explanations.
Gemma 3 · 1B
A tiny, text-only starting point for simple rewrites and short questions. Expect limited answers.
Qwen3 · 1.7B
Try short drafts, summaries, and everyday questions on a smaller laptop.
Qwen3 · 4B
Work through questions, brainstorm ideas, and try local reasoning without a large download.
Phi-4 Mini · 3.8B
A compact option for math, logic, and explaining small pieces of code.
Gemma 3 · 4B
Ask questions about images, draft text, and summarize with a compact multimodal model.
Qwen3 · 8B
A step up for reasoning, coding questions, and more detailed everyday conversations.
Catalog checked October 15, 2026. Model details: Ollama model library. Memory guidance: LM Studio system requirements. Check the model listing and your runtime’s fit estimate before downloading.
From download to first reply
Get running on your computer.
Pick your setup. These short guides use desktop apps and everyday hardware.
Windows PCYour first local chatA familiar desktop app is all you need to get started.
- Install LM Studio from its official website, then open the app.
- In Discover, pick a small instruction-tuned model from the recommendations above. Choose a 4-bit GGUF download when available.
- Go to Chat, open the model picker, and load your downloaded model.
- Try a short request first. If loading fails or replies crawl, choose a smaller model and keep the context length modest.
LM Studio recommends 16 GB of RAM. A supported GPU can help; on CPU, small models are the practical starting point. Check the app’s requirements for your processor.
Windows + NVIDIAPut your graphics card to workUse the GPU you already own, with a normal desktop installer.
- Check your GPU model and dedicated GPU memory in Task Manager → Performance → GPU. This is separate from your PC’s RAM.
- Update your NVIDIA driver and check your card against Ollama’s supported GPU list.
- Install the Ollama desktop app for Windows. In the app, select and download a local model that fits your memory estimate above.
- Send a short message. If memory runs out, close other GPU-heavy apps, reduce context length in settings, or switch to a smaller model.
VRAM matters more than the “RTX” name for fitting a model. System RAM and VRAM are separate limits; fitting partly in RAM can be slower. Choose a downloaded local model rather than a cloud option.
Apple silicon MacMake the most of unified memoryA straightforward route for M-series MacBooks, iMacs, and Mac minis.
- Open About This Mac to check your M-series chip and memory. LM Studio requires macOS 14 or newer.
- Install LM Studio for Apple silicon and open Discover.
- Download a small model from the recommendations above. GGUF works on Mac; MLX is another supported format for Apple silicon.
- Load the model in Chat and try a short prompt. Close memory-heavy apps and reduce context length if your Mac becomes sluggish.
Your CPU and GPU share unified memory, and macOS needs some too. An 8 GB Mac should start with small models; 16 GB or more gives you more room. Intel Macs are not supported by LM Studio.
Download first, then chat offline. LM Studio runs downloaded models on your device; finding models, downloads, and updates still need an internet connection. How offline mode works