Hardware, drawn to scale
Two numbers decide what a computer can do with AI. How much memory it can give a model, and how fast it can read that memory.
Every model, drawn to scale
Each block’s area is the memory a model needs. The dashed square is what your machine can give it. If the block fits inside, the model runs.
4 of 12 run well on this machine.
- Llama 3.2 3B22–37 tokens/s, of 10.7 GB available. Opens in the checker.
- Gemma 4 E4B19–31 tokens/s, of 10.7 GB available. Opens in the checker.
- Qwen3 8B10–16 tokens/s, of 10.7 GB available. Opens in the checker.
- Llama 3.1 8B10–17 tokens/s, of 10.7 GB available. Opens in the checker.
- Gemma 3 12BToo big, of 10.7 GB available. Opens in the checker.
- gpt-oss 20BToo big, of 10.7 GB available. Opens in the checker.
- Mistral Small 3.2Too big, of 10.7 GB available. Opens in the checker.
- Gemma 4 26B-A4BToo big, of 10.7 GB available. Opens in the checker.
- Qwen 3.6 27BToo big, of 10.7 GB available. Opens in the checker.
- Gemma 4 31BToo big, of 10.7 GB available. Opens in the checker.
- Qwen 3.6 35B-A3BToo big, of 10.7 GB available. Opens in the checker.
- Llama 3.3 70BToo big, of 10.7 GB available. Opens in the checker.
Choose a model to open it in the checker with these settings.
- Published: From the model's published architecture.Model sizes: each model’s published architectureChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.Usable memory: our estimate, until measuredChecked: not yet verified
What each machine gets you
Memory decides what you can run. Memory speed decides how fast it writes. Every model here at 4-bit with 8K tokens of conversation.
Mac
- Mac mini M4, 16 GBIn my setupUsable by models10.7 GBMemory speed120 GB/sBiggest that runs wellQwen3 8B10–16 tokens/s · 4 of 12 models
- MacBook Air M4, 24 GBUsable by models16.1 GBMemory speed120 GB/sBiggest that runs wellgpt-oss 20B27–45 tokens/s · 6 of 12 models
- MacBook Pro M4 Pro, 24 GBUsable by models16.1 GBMemory speed273 GB/sBiggest that runs wellgpt-oss 20B61–102 tokens/s · 6 of 12 models
- MacBook Pro M4 Pro, 48 GBUsable by models36 GBMemory speed273 GB/sBiggest that runs wellQwen 3.6 35B-A3B65–108 tokens/s · 11 of 12 models
- MacBook Pro M4 Max, 64 GBUsable by models48 GBMemory speed546 GB/sBiggest that runs wellQwen 3.6 35B-A3B129–215 tokens/s · 11 of 12 models
PC with NVIDIA
- RTX 4060Usable by models7.2 GBMemory speed272 GB/sBiggest that runs wellGemma 4 E4B43–71 tokens/s · 2 of 12 models
- RTX 4070Usable by models11.2 GBMemory speed504 GB/sBiggest that runs wellQwen3 8B41–68 tokens/s · 4 of 12 models
- RTX 3090Usable by models23 GBMemory speed936 GB/sBiggest that runs wellQwen 3.6 27B25–42 tokens/s · 9 of 12 models
- RTX 4090In my setupUsable by models23 GBMemory speed1008 GB/sBiggest that runs wellQwen 3.6 27B27–45 tokens/s · 9 of 12 models
- RTX 5090Usable by models31 GBMemory speed1792 GB/sBiggest that runs wellQwen 3.6 35B-A3B424–706 tokens/s · 11 of 12 models
- Spec sheet: From the manufacturer's published specifications.Memory and memory speed: Apple and NVIDIA spec sheetsChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.Usable memory and speeds: our estimates, until measuredChecked: not yet verified
Mac or PC?
They get to the same place by different routes.
A Mac
Has Unified memoryOn Apple Silicon Macs, one pool of memory shared by the processor and the graphics chip.Read the full entry: one pool shared by everything, so a 64 GB Mac can give a model about 48 GB. That fits models no single consumer graphics card can hold.
- Quiet, efficient, and the same setup on a laptop or desktop.
- Memory speed depends on the chip: 120 GB/s for the M4, 273 for the M4 Pro, 546 for the M4 Max.
- The model shares memory with your apps, so close big ones first.
A PC with an NVIDIA card
Runs models from the card’s own memory, its VRAMThe memory on a graphics card. Many times faster than a PC’s main memory, and usually the limit on what fits.Read the full entry. That memory is very fast, up to 1792 GB/s, but there is only 8 to 32 GB of it.
- The fastest way to run any model that fits on the card.
- Models that don’t fit Spilling overWhen a model is too big for the graphics card, part of it runs from the PC’s slower main memory.Read the full entry into main memory and slow right down.
- Loud and power-hungry under load; better as a desktop that stays put.
If you are choosing a new one
What matters for running models, in order.
- Get more memory than you think. It decides which models you can run at all, and on a Mac you can’t add more later. 16 GB runs good small models; 32 GB or more opens up the mid-size ones that feel close to the cloud.
- Then look at memory speed. Two machines with the same memory can write at very different speeds. An M4 Pro writes about twice as fast as an M4 with the same model.
- On a PC, the card’s memory beats its rank. An older card with 24 GB, like the RTX 3090, runs models a newer 12 GB card can’t fit at all.
- Processor and storage matter least. A fast drive makes models load sooner, but it doesn’t change how fast they write.
Find out what you have
On a Mac
Open the Apple menu and choose About This Mac. Note your Chip, such as Apple M4 Pro, and Memory, such as 24 GB.
On Windows
Press Ctrl Shift Esc to open Task Manager, then choose Performance and GPU. The card’s name is at the top right, and Dedicated GPU memory is how much it has.