Mac vs PC for local LLMs

Mac Studio bandwidth from Apple's spec page; GPUs from TechPowerUp. Mac unified memory is shared with macOS, so treat fit results as optimistic. Mac prices: check price (we couldn't map apple.com prices to configs reliably).

Spec	Apple Mac Studio M5 Max (40-core GPU, 64 GB)	Apple Mac Studio M5 Ultra (64-core GPU, 96 GB)	GeForce RTX 3090	GeForce RTX 5090
Memory	64 GB Unified LPDDR	96 GB Unified LPDDR	24 GB GDDR6X	32 GB GDDR7
Bandwidth	614 GB/s	1200 GB/s	936.2 GB/s	1790 GB/s
Low price	check price	check price	low $1,889, avg $2,404 (used) source, 2026-10-15	low $6,500, avg $8,757 (new) source, 2026-10-15
Qwen2.5-Coder-32B-Instruct estimate	~24 tok/s	~46 tok/s	~36 tok/s	~69 tok/s
$ per tok/s (Qwen2.5-Coder-32B-Instruct)	check price	check price	$52	$94
Models that fit	25 of 27	27 of 27	22 of 27	22 of 27
Largest that fits	Qwen2.5-72B-Instruct	GLM-4.5-Air	Qwen2.5-Coder-32B-Instruct	Qwen2.5-Coder-32B-Instruct

Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.

What each one runs, model by model

Model (Q4_K_M)	File	Apple Mac Studio M5 Max (40-core GPU, 64 GB) estimate	Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) estimate	GeForce RTX 3090 estimate	GeForce RTX 5090 estimate
GLM-4.5-Air	73.5 GB	doesn't fit	~114 tok/s	doesn't fit	doesn't fit
gpt-oss-120b	62.84 GB	doesn't fit	~332 tok/s	doesn't fit	doesn't fit
Qwen2.5-72B-Instruct	47.42 GB	~10 tok/s	~19 tok/s	doesn't fit	doesn't fit
Llama-3.3-70B-Instruct	42.52 GB	~11 tok/s	~21 tok/s	doesn't fit	doesn't fit
DeepSeek-R1-Distill-Llama-70B	42.52 GB	~11 tok/s	~21 tok/s	doesn't fit	doesn't fit
Qwen2.5-Coder-32B-Instruct	19.85 GB	~24 tok/s	~46 tok/s	~36 tok/s	~69 tok/s
DeepSeek-R1-Distill-Qwen-32B	19.85 GB	~24 tok/s	~46 tok/s	~36 tok/s	~69 tok/s
Qwen3-32B	19.76 GB	~24 tok/s	~46 tok/s	~36 tok/s	~69 tok/s
Qwen3-30B-A3B	18.63 GB	~232 tok/s	~453 tok/s	~353 tok/s	~676 tok/s
gemma-3-27b-it	16.55 GB	~28 tok/s	~55 tok/s	~43 tok/s	~82 tok/s
Mistral-Small-3.2-24B-Instruct-2506	14.33 GB	~33 tok/s	~64 tok/s	~50 tok/s	~95 tok/s
Devstral-Small-2505	14.33 GB	~33 tok/s	~64 tok/s	~50 tok/s	~95 tok/s
gpt-oss-20b	11.67 GB	~232 tok/s	~454 tok/s	~354 tok/s	~677 tok/s
Qwen3-14B	9 GB	~52 tok/s	~101 tok/s	~79 tok/s	~151 tok/s
Qwen2.5-Coder-14B-Instruct	8.99 GB	~52 tok/s	~101 tok/s	~79 tok/s	~151 tok/s
DeepSeek-R1-Distill-Qwen-14B	8.99 GB	~52 tok/s	~101 tok/s	~79 tok/s	~151 tok/s
Mistral-Nemo-Instruct-2407	7.48 GB	~62 tok/s	~122 tok/s	~95 tok/s	~182 tok/s
gemma-3-12b-it	7.3 GB	~64 tok/s	~125 tok/s	~97 tok/s	~186 tok/s
NVIDIA-Nemotron-Nano-9B-v2	6.53 GB	~71 tok/s	~140 tok/s	~109 tok/s	~208 tok/s
Qwen3-8B	5.03 GB	~93 tok/s	~181 tok/s	~141 tok/s	~270 tok/s
Llama-3.1-8B-Instruct	4.92 GB	~95 tok/s	~185 tok/s	~145 tok/s	~277 tok/s
Qwen2.5-Coder-7B-Instruct	4.68 GB	~100 tok/s	~195 tok/s	~152 tok/s	~291 tok/s
Mistral-7B-Instruct-v0.3	4.37 GB	~107 tok/s	~209 tok/s	~163 tok/s	~311 tok/s
Qwen3-4B	2.5 GB	~187 tok/s	~365 tok/s	~285 tok/s	~544 tok/s
gemma-3-4b-it	2.49 GB	~187 tok/s	~366 tok/s	~286 tok/s	~546 tok/s
Llama-3.2-3B-Instruct	2.02 GB	~231 tok/s	~451 tok/s	~352 tok/s	~673 tok/s
Qwen3-0.6B	0.48 GB	~972 tok/s	~1900 tok/s	~1482 tok/s	~2834 tok/s

File sizes are Q4_K_M GGUFs from Hugging Face (linked on each model page), accessed 2026-10-15. A model fits when its file plus 1.5 GB of headroom is at most the card's memory.

Speed calculator estimate

GPU Count Model

Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works

Questions people ask

Which is faster on Qwen2.5-Coder-32B-Instruct?

On Qwen2.5-Coder-32B-Instruct (19.85 GB at Q4_K_M), the largest model that fits all of them: Mac Studio M5 Max about 24 tok/s, Mac Studio M5 Ultra about 46 tok/s, GeForce RTX 3090 about 36 tok/s, GeForce RTX 5090 about 69 tok/s. Speed tracks memory bandwidth: Mac Studio M5 Max 614 GB/s, Mac Studio M5 Ultra 1200 GB/s, GeForce RTX 3090 936.2 GB/s, GeForce RTX 5090 1790 GB/s.

Which is cheaper per token per second?

Lowest tracked price divided by the estimated speed on Qwen2.5-Coder-32B-Instruct: Mac Studio M5 Max no verified price, Mac Studio M5 Ultra no verified price, GeForce RTX 3090 $52 per tok/s, GeForce RTX 5090 $94 per tok/s. Those prices are Mac Studio M5 Max no verified price, Mac Studio M5 Ultra no verified price, GeForce RTX 3090 $1,889 (used), GeForce RTX 5090 $6,500 (new), dated snapshots that move daily.

What fits on each of them?

Mac Studio M5 Max (64 GB): 25 of our 27 sized models, largest Qwen2.5-72B-Instruct at 47.42 GB; Mac Studio M5 Ultra (96 GB): 27 of our 27 sized models, largest GLM-4.5-Air at 73.5 GB; GeForce RTX 3090 (24 GB): 22 of our 27 sized models, largest Qwen2.5-Coder-32B-Instruct at 19.85 GB; GeForce RTX 5090 (32 GB): 22 of our 27 sized models, largest Qwen2.5-Coder-32B-Instruct at 19.85 GB. Unified memory on the Apple and Framework systems is shared with the OS, so treat those fit results as optimistic.

Which of these runs the largest model?

Mac Studio M5 Ultra with 96 GB: it fits GLM-4.5-Air (73.5 GB at Q4_K_M), an estimated 114 tok/s. The smallest memory here, 24 GB on the GeForce RTX 3090, tops out at Qwen2.5-Coder-32B-Instruct (19.85 GB). Unified memory on the Apple and Framework systems is shared with the OS, so treat those fit results as optimistic.