What is the biggest model the RTX 4090 can run?

Qwen3 32B, which needs about 22 GB, on the 24 GB version at about 28 tok/s.

What is the biggest model the Apple M4 Max can run?

gpt-oss 120B, which needs about 67 GB, on the 128 GB version at about 66 tok/s.

RTX 4090 vs Apple M4 Max for local AI

The Apple M4 Max fits more: up to gpt-oss 120B, against Qwen3 32B on the RTX 4090. The RTX 4090 generates about 85% faster on any model both can run, because it has 1008 GB/s of memory bandwidth to the Apple M4 Max's 546 GB/s.

The numbers that matter

Memory and bandwidth

Memory decides which models fit. Bandwidth decides how fast they generate, because every token reads the model's active weights once.

	RTX 4090	Apple M4 Max
Memory	24 GB VRAM	36 or 48 or 64 or 128 GB unified memory
Usable for models	All 24 GB	About 96 GB of 128 GB
Bandwidth	1008 GB/s	546 GB/s
Biggest model that fits	Qwen3 32B	gpt-oss 120B
Best everyday pick	Qwen3 32B, ~28 tok/s	gpt-oss 120B, ~66 tok/s

Model by model

Every model, side by side

Estimated generation speed at the largest memory size each was sold in (24 GB and 128 GB).

Model	RTX 4090	Apple M4 Max
Qwen3 4B About 4 GB	~220 tok/s Runs great	~120 tok/s Runs great
Qwen3 8B About 7 GB	~110 tok/s Runs great	~58 tok/s Runs great
Gemma 3 12B About 10 GB	~68 tok/s Runs great	~37 tok/s Runs great
Qwen3 14B About 11 GB	~60 tok/s Runs great	~32 tok/s Runs great
gpt-oss 20B About 16 GB	~180 tok/s Runs great	~96 tok/s Runs great
Gemma 3 27B About 19 GB	~33 tok/s Runs great	~18 tok/s Usable
Qwen3 30B-A3B About 21 GB	~180 tok/s Runs great	~96 tok/s Runs great
Qwen3 32B About 22 GB	~28 tok/s Usable	~15 tok/s Usable
Llama 3.3 70B About 45 GB	Needs 45 GB	~7 tok/s Slow
gpt-oss 120B About 67 GB	Needs 67 GB	~66 tok/s Runs great
Qwen3 235B-A22B About 144 GB	Needs 144 GB	Needs 144 GB
DeepSeek R1 671B About 406 GB	Needs 406 GB	Needs 406 GB

Which should you get?

It depends. The Apple M4 Max fits bigger models; the RTX 4090 runs the models both fit faster. Price matters as much as either number: check used prices on the builds page, where every parts list has a measured or estimated speed.

How we estimate

Both columns use the same formula: memory bandwidth divided by the weights read per token, calibrated on our own rack, with Ollama's default 4-bit sizes. Real results are usually within about a third. Measured either one? Send us the number.

Compare

Other matchups

Every comparison