Free local AI tool

Local LLM Picker

The Local LLM Picker tells you which open AI models you can run on your own computer. Enter your GPU VRAM, Mac unified memory or system RAM and a use case, and it ranks 44 current Ollama and LM Studio models by fit, with the best pick, its speed and the exact command to run it.

Find the best local model for your machine

Pick your hardware, memory and what you want the model to do. Results update instantly and nothing leaves your browser.

Hardware
Common sizes
Use case

Best pick

Gemma 4 12B

8.02 GB download · fits in about 11.5 GB usable · fast speed · 256K context · Apache 2.0

ollama run gemma4:12b

All chat & writing models

13 of 24 fit · data checked November 12, 2026

Local models ranked for your hardware
ModelFitSizeSpeedContextLicenseCommand
Gemma 4 12BFits8.02 GBFast256KApache 2.0ollama run gemma4:12b
Ministral 3 14BFits9.08 GBFast256KApache 2.0ollama run ministral-3:14b
Qwen3.5 9BFits6.55 GBFast256KApache 2.0ollama run qwen3.5:9b
Ministral 3 8BFits6.02 GBFast256KApache 2.0ollama run ministral-3:8b
Gemma 4 E4BFits6.58 GBFast128KApache 2.0ollama run gemma4:e4b
Qwen3.5 4BFits3.32 GBFast256KApache 2.0ollama run qwen3.5:4b
Gemma 4 E2BFits4.59 GBFast128KApache 2.0ollama run gemma4:e2b
Llama 3.1 8BFits4.92 GBFast128KLlama 3.1 Communityollama run llama3.1:8b
Qwen3.5 2BFits2.68 GBFast256KApache 2.0ollama run qwen3.5:2b
Phi-4 Mini 3.8BFits2.49 GBFast128KMITollama run phi4-mini:3.8b
Llama 3.2 3BFits2.02 GBFast128KLlama 3.2 Communityollama run llama3.2:3b
Qwen3.5 0.8BFits1.32 GBFast256KApache 2.0ollama run qwen3.5:0.8b
Gemma 3 1BFits0.81 GBFast32KGemma Terms of Useollama run gemma3:1b
gpt-oss 20BToo big13.79 GB-128KApache 2.0ollama run gpt-oss:20b
Mistral Small 3.2 24BToo big15.18 GB-128KApache 2.0ollama run mistral-small3.2:24b
Qwen3.5 27BToo big17.48 GB-256KApache 2.0ollama run qwen3.5:27b
Qwen3.8 27BToo big17.74 GB-256KApache 2.0ollama run qwen3.8:27b
Gemma 4 26B (MoE)Too big18.73 GB-256KApache 2.0ollama run gemma4:26b
Gemma 4 31BToo big20.39 GB-256KApache 2.0ollama run gemma4:31b
Qwen3.5 35B (MoE)Too big22.07 GB-256KApache 2.0ollama run qwen3.5:35b
Qwen3.6 35B (MoE)Too big22.62 GB-256KApache 2.0ollama run qwen3.6:35b
Llama 3.3 70BToo big42.52 GB-128KLlama 3.3 Communityollama run llama3.3:70b
gpt-oss 120BToo big65.37 GB-128KApache 2.0ollama run gpt-oss:120b
Qwen3.5 122B (MoE)Too big81.37 GB-256KApache 2.0ollama run qwen3.5:122b

Sizes are Ollama's default downloads (mostly 4-bit Q4_K_M). "Fits" leaves about 15% plus 0.5 GB for a few thousand tokens of context. Speed is a rough tier from the active weights and typical memory bandwidth, not a benchmark.

How to choose a local LLM

  1. Step 1

    Pick your hardware

    Choose NVIDIA or AMD GPU, Apple Silicon Mac, or CPU only.

  2. Step 2

    Enter your memory

    Type your GPU VRAM, Mac unified memory or system RAM in GB, or tap a common size.

  3. Step 3

    Choose a use case

    Chat, coding, reasoning, images or embeddings. The table ranks every model that fits.

  4. Step 4

    Run it

    Copy the ollama run command into a terminal, or search the model name in LM Studio.

Frequently asked questions

What is the best local LLM to run?

It depends on your memory. An 8 GB Mac runs a 4B model such as Qwen3.5 4B well. A 16 GB GPU fits Gemma 4 12B comfortably and gpt-oss 20B just fits. A 24 GB GPU fits 27B models such as Qwen3.8 27B. A 96 GB Mac runs Qwen3 Coder Next, with gpt-oss 120B as a tight fit.

How much RAM or VRAM do I need to run an LLM locally?

Roughly the download size plus 15% and half a gigabyte for context. A 4-bit model needs about 0.6 GB per billion parameters, so a 9B model wants about 8 GB and a 27B model about 21 GB. Longer chats need more for the KV cache.

How much of a Mac's memory can a local model use?

By default macOS lets the GPU use about two-thirds of unified memory on Macs with 36 GB or less, and about three-quarters on larger Macs. A 16 GB Mac has about 10.7 GB for a model; a 64 GB Mac has about 48 GB.

Should I use Ollama or LM Studio?

Both run the same open models. Ollama is a command-line tool and local API (ollama run qwen3.5:9b), good for developers and apps. LM Studio is a desktop app with a model search and chat window, good if you prefer clicking. This picker gives you the Ollama command and the name to search in LM Studio.

Why are mixture-of-experts (MoE) models faster?

A MoE model such as Qwen3.6 35B or gpt-oss 20B stores all its weights but only reads a few billion of them for each token. It needs the memory of a big model but generates text at the speed of a small one, which suits Macs and laptops.

Model sizes, context windows and licenses were checked against the Ollama library on November 12, 2026. Rankings are an editorial guide, not benchmark results. Ollama and LM Studio are products of their respective owners; AI Wins is not affiliated with them.

Related AI tools

LLM VRAM Calculator

Estimate memory for any model size, quantization and context length.

Open tool

LLM Token Counter

Count tokens in your prompt before you hit a context limit.

Open tool

AI Model Context Window Comparison

Compare how much text each AI model can read at once.

Open tool

AI API Pricing Calculator

Compare the cost of hosted models when local is not enough.

Open tool

New open models land every week. Get the ones worth trying in the free daily AI Wins digest.

Get the free digest

Get the free AI Wins digest

Only positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.