Chat · Google
Runs great
Gemma 4 E4B
Best small model for everyday questions and writing.
20 words/ssmoothNeeds ~7.4 GB (6.6 GB download)
Pick your Mac or PC and see which free AI models run on it, how fast, and the exact command to start. No signup.
Your M4 16GB
Best pick for chat: Gemma 4 12B. For coding: Qwen2.5 Coder 7B.
Best small model for everyday questions and writing.
20 words/ssmoothThinks step by step before answering. Good for maths and logic.
12 words/ssmoothLight code helper for autocomplete on smaller machines.
14 words/ssmoothTiny and quick. Runs on almost anything, phones included.
32 words/sfaster than you can readSmall all-rounder that reads images too.
19 words/ssmoothMakes full songs with vocals from lyrics.
about 3 minper 1 minute songUse the ACE-Step app
Sharp, modern images and photo editing from a small model.
about 34 sper imageUse Draw Things (Mac) or ComfyUI
Strong at maths and reasoning for its size.
7 words/sabout reading speedSweet spot for 16 GB machines. Good writing and reasoning.
8 words/sabout reading speedClassic image model with thousands of community styles.
about 18 sper imageUse Draw Things (Mac) or ComfyUI
Clone a voice from a short sample. Very natural.
about 52 sper minute of speechUse Voicebox, or the Qwen3-TTS repo
Strong everyday helper, 256K context, sees images.
10 words/sabout reading speedReads photos, screenshots and documents.
11 words/sabout reading speedSpeech to text in 90+ languages. Free, private transcripts.
about 1 minper 10 minutes of audioUse MacWhisper, or whisper.cpp
Natural text to speech from a tiny 82M model.
about 11 sper minute of speechpip install kokoro
OpenAI's open model. Smart for its size and fast because only part of it runs per word.
13 words/ssmoothGreat photo-real scenes in a few steps. Weak at writing text in images.
about 2 minper imageUse ComfyUI, or mflux on a Mac
Text or image to video on a single consumer GPU.
about 27 minper 4 second clipUse ComfyUI
Frontier class. Needs a 512 GB Mac Studio or a server.
Near frontier quality at home, if you have 96 GB or more.
OpenAI's big open model. Fast for its size on high-memory machines.
Older but solid large model. Needs serious memory.
Best open video quality, but heavy.
Short clips with real camera motion. Detail is soft at low settings.
Fast coding and agent model, mixture of experts.
Very fast for its size. Great general assistant.
One of the best models that fits a 24 GB GPU or 32 GB Mac.
Serious reasoning on a 24 GB GPU or 32 GB Mac.
Top local coding model for 24 GB GPUs and 32 GB Macs.
Newest Qwen, tuned for agentic coding and long tasks.
Mixture of experts: big-model quality at small-model speed.
The models you would gain with more memory or a stronger chip.
Unlocks 4 more: Gemma 4 26B A4B, Qwen3 Coder 30B, Z-Image Turbo 6B, Wan 2.2 5B.
Unlocks 11 more: Gemma 4 26B A4B, Qwen 3.5 35B A3B, Gemma 4 31B, Qwen3 Coder 30B and more.
Unlocks 14 more: Gemma 4 26B A4B, Qwen 3.5 35B A3B, Gemma 4 31B, Llama 3.3 70B and more.
Each model needs a certain amount of memory: its download size plus room for the conversation. If that fits in your graphics memory (or a Mac's unified memory), it runs fast. If it only fits with help from normal RAM, it runs slowly. Speed is estimated from your memory bandwidth, because generating text is mostly limited by how fast the chip can read the model.
Ollama is a free app for Mac, Windows and Linux that downloads and runs these models with one command. Install it from ollama.com, open a terminal, and paste the command shown on any model card. Prefer buttons? LM Studio does the same with a normal app window.
Yes. Once downloaded, the model runs entirely on your computer. Your prompts, files and photos never leave the machine, and it works offline.
Macs share one big pool of memory with the GPU, so a 64 GB or 128 GB Mac can load models no gaming card can. NVIDIA cards are faster for models that fit in their memory, and best for image and video tools. For the biggest models per dollar, a high-memory Mac or a Ryzen AI Max box wins.
Longer conversations, bigger context windows, other apps and different quantizations all change speed and memory. Treat the numbers here as a good first guide, not a lab measurement.