Check what you have
About 2 minutes
Find your chip or graphics card, and its memory. The checker above does the rest.
Run AI models on your own Mac or PC
On your , runs comfortably.
Qwen3 8B runs comfortably. Like a fast typist: about 41–68 tokens a second. It needs 6.8 GB of the 11.2 GB your graphics card has free.
Assuming precision and of conversation.
The model 5.0 GB
8.19 billion numbers at 4.85 bits each. Lower precision makes it smaller, with slightly duller answers.
Conversation memory 1.2 GB
The model keeps notes on everything said so far, about a short report here. Longer chats need bigger notes.
Why the card? 4.4 GB free
A graphics card’s own memory is many times faster than the PC’s main memory. A model that fits on the card flies; one that spills over crawls.
Estimated speed Estimate: Calculated from a stated rule of thumb. Real results vary.
41–68
tokens per second. A token is about three-quarters of a word.
ollama run qwen3:8bRace it against
Qwen3 8B on your machineAbout 55 a second
Claude Sonnet 5.5 in the cloudAbout 139 a second
Writing speed only, not thinking time. Yours is our estimate; the cloud’s is measured by Artificial Analysis.
- Published: From the model's published architecture.Model sizes: each model’s published architectureChecked: not yet verified
- Spec sheet: From the manufacturer's published specifications.Memory and bandwidth: Apple and NVIDIA spec sheetsChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.Usable memory and speed: our estimates, until measuredChecked: not yet verified
- Tested by others: Measured by a named independent tester, not by us.Cloud writing speeds: Artificial AnalysisChecked: 31 Oct 2026
Everything, on your machine
PC with an RTX 4070, at 4-bit with 8K tokens of conversation. Pick one to see it above.
On your , runs comfortably.
Qwen3 8B runs comfortably. About reading speed: about 10–16 tokens a second. It needs 6.8 GB of the 10.7 GB your Mac can give it.
Assuming precision and of conversation.
The model 5.0 GB
8.19 billion numbers at 4.85 bits each. Lower precision makes it smaller, with slightly duller answers.
Conversation memory 1.2 GB
The model keeps notes on everything said so far, about a short report here. Longer chats need bigger notes.
Why not all 16 GB? 3.9 GB free
macOS lets models use about 67% of your Mac’s memory, keeping the rest for everything else.
Estimated speed Estimate: Calculated from a stated rule of thumb. Real results vary.
10–16
tokens per second. A token is about three-quarters of a word.
ollama run qwen3:8bRace it against
Qwen3 8B on your machineAbout 13 a second
Claude Sonnet 5.5 in the cloudAbout 139 a second
Writing speed only, not thinking time. Yours is our estimate; the cloud’s is measured by Artificial Analysis.
- Published: From the model's published architecture.Model sizes: each model’s published architectureChecked: not yet verified
- Spec sheet: From the manufacturer's published specifications.Memory and bandwidth: Apple and NVIDIA spec sheetsChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.Usable memory and speed: our estimates, until measuredChecked: not yet verified
- Tested by others: Measured by a named independent tester, not by us.Cloud writing speeds: Artificial AnalysisChecked: 31 Oct 2026
Everything, on your machine
Mac mini M4 with 16 GB, at 4-bit with 8K tokens of conversation. Pick one to see it above.
Fits is a cliff, not a slope
Bigger models answer better, so you want the biggest one that fits. Drag through them: speed eases down as they grow, then falls the moment one outgrows fast memory.
On a Mac mini M4 with 16 GB, with 10.7 GB of fast memory for models. Change machine
Qwen3 8B needs 6.8 GB
10–16 tokens a second Estimate: Calculated from a stated rule of thumb. Real results vary.
About reading speed
Every model in our checker, at 4-bit with 8K tokens of conversation. Drag, or use the arrow keys.
- The bar: what the model needs
- The line: a plain model of any size
- A model in our checker
- Mixture of experts: reads only part of itself per word, so it beats the line
- Spec sheet: From the manufacturer's published specifications.Memory and bandwidth: Apple and NVIDIA spec sheetsChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.Speeds: our estimates, until measuredChecked: not yet verified
Your first model in about 30 minutes
No coding, no accounts. One free app, one download, and a model that never leaves your computer.
Start hereInstall Ollama
About 5 minutes
A free app that runs models in the background.
Download a model and chat
About 5 to 20 minutes, depending on your internet
Pick one sized for your machine and ask it something.
Local or cloud? Usually both.
Local is best when
- What you type is private: contracts, health, customer data.
- You want it to work offline, or never hit a usage limit.
- The task is routine: summarising, rewriting, sorting, quick code.
Claude, ChatGPT or Grok are best when
- You need the strongest reasoning or the newest knowledge.
- The job is long and hard: big codebases, research, deep analysis.
- You are on a phone or a machine without much memory.
How I run it
One real setup: a gaming PC that serves AI to the whole house, and Macs that use it over Wi-Fi.
- Alienware desktopRTX 4090 with 24 GB, 64 GB RAM. Always on, serving models to the house.
- Mac mini M416 GB. Runs small models itself; borrows the PC for big ones.
- MacBooksDay-to-day machines that talk to the PC over Wi-Fi.