Apple · M3 Ultra · Desktop

Best AI Models for Mac Studio (M3 Ultra, 512GB)

Runs 27–31B and 120B MoE models well; dense 70B fits but is slow at 12 tok/s.

512 GB unified508.5 GB for models819 GB/s66 TFLOPS FP16270 W2025

Memory

96 GB$3,999 256 GB$7,099 512 GB$9,499

What LLMs can the Mac Studio (M3 Ultra, 512GB) run?

#	Model	Speed	Quant · Memory
1	DeepSeek V4.1 Flash 1017DeepSeekHow to run → Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
2	MiMo V2.6 Flash RLXiaomi Q4_K_M · 186.5 GB · 39.5 t/s	39.5 tok/s	Q4_K_M · 186.5 GB
3	GLM 5.3Z.ai Q4_K_M · 453.8 GB · 16.7 t/s	16.7 tok/s	Q4_K_M · 453.8 GB
4	DeepSeek V4.1 FlashDeepSeek Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

#	Model	Speed	Quant · Memory
1	DeepSeek V4.1 Flash 1017DeepSeekHow to run → Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
2	GLM 5.3Z.ai Q4_K_M · 453.8 GB · 16.7 t/s	16.7 tok/s	Q4_K_M · 453.8 GB
3	Muse Spark Open 816B A42BMeta Q4_0 · 461.5 GB · 20.0 t/s	20.0 tok/s	Q4_0 · 461.5 GB
4	MiMo V2.6 Flash RLXiaomi Q4_K_M · 186.5 GB · 39.5 t/s	39.5 tok/s	Q4_K_M · 186.5 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

#	Model	Speed	Quant · Memory
1	DeepSeek V4.1 Flash 1017DeepSeekHow to run → Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
2	DeepSeek V4.1 FlashDeepSeek Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
3	GLM 5.3Z.ai Q4_K_M · 453.8 GB · 16.7 t/s	16.7 tok/s	Q4_K_M · 453.8 GB
4	Muse Spark Open 816B A42BMeta Q4_0 · 461.5 GB · 20.0 t/s	20.0 tok/s	Q4_0 · 461.5 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

#	Model	Speed	Quant · Memory
1	MiMo V2.6 Flash RLXiaomiHow to run → Q4_K_M · 186.5 GB · 39.5 t/s	39.5 tok/s	Q4_K_M · 186.5 GB
2	DeepSeek V4 Flash 0731DeepSeek Q4_K_M · 183.0 GB · 26.8 t/s	26.8 tok/s	Q4_K_M · 183.0 GB
3	GLM 5.3 FlashZ.ai Q4_K_M · 205.2 GB · 35.3 t/s	35.3 tok/s	Q4_K_M · 205.2 GB
4	Qwen3.8 27BAlibaba Q4_K_M · 17.9 GB · 32.9 t/s	32.9 tok/s	Q4_K_M · 17.9 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

#	Model	Speed ↓	Quant · Memory
1	Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 176 t/s	176 tok/s	Q4_K_M · 3.4 GB
2	GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 107 t/s	107 tok/s	Q4_K_M · 13.1 GB
3	Inkling SmallThinking Machines Q4_K_M · 162.7 GB · 98.0 t/s	98.0 tok/s	Q4_K_M · 162.7 GB
4	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 92.7 t/s	92.7 tok/s	Q4_K_M · 16.5 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

#	Model	Speed	Quant · Memory
1	DeepSeek V4.1 Flash 1017DeepSeekHow to run → Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
2	DeepSeek V4.1 FlashDeepSeek Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
3	Muse Spark Open 816B A42BMeta Q4_0 · 461.5 GB · 20.0 t/s	20.0 tok/s	Q4_0 · 461.5 GB
4	GLM 5.3 FlashZ.ai Q4_K_M · 205.2 GB · 35.3 t/s	35.3 tok/s	Q4_K_M · 205.2 GB
5	NVIDIA Nemotron 3.5 Super VL 120B A12B BF16NVIDIA Q4_K_M · 87.7 GB · 42.7 t/s	42.7 tok/s	Q4_K_M · 87.7 GB

#	Model	Speed	Quant · Memory
1	DeepSeek V4.1 Flash 1017DeepSeekHow to run → Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
2	GLM 5.3Z.ai Q4_K_M · 453.8 GB · 16.7 t/s	16.7 tok/s	Q4_K_M · 453.8 GB
3	GLM 5.2Z.ai Q4_K_M · 453.8 GB · 16.7 t/s	16.7 tok/s	Q4_K_M · 453.8 GB
4	DeepSeek V4.1 FlashDeepSeek Q4_K_M · 458.4 GB · 40.4 t/s	40.4 tok/s	Q4_K_M · 458.4 GB
5	Hy3Tencent Q4_K_M · 182.3 GB · 26.3 t/s	26.3 tok/s	Q4_K_M · 182.3 GB

See all 439 models on the Mac Studio (M3 Ultra, 512GB) →Bars = traction for models without a Score yet

About the Mac Studio (M3 Ultra, 512GB)

The Mac Studio (M3 Ultra, 512GB) runs 27–31B and 120B MoE models well. macOS keeps about 3.5 GB, leaving 508.5 GB for models. Prompt reading is its weak spot: a 4K-token prompt waits 10 s before the first word on DeepSeek V4.1 Flash 1017. At its $9,499 MSRP it gives 8.1 tok/s per $1,000, 6th of the 6 machines compared here. M3 Ultra generation, released 2025.

For models508.5 GBmacOS keeps 3.5 GB

ChipM3 Ultrareleased 2025

Mac Studio (M3 Ultra, 512GB) strengths and weaknesses

Strengths

Runs 120B MoE models: GPT OSS 120B at 75.5 tok/s in 70.8 GB
Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 490.6 GB free

Weaknesses

Slow prompt reading: A 32K-token prompt on DeepSeek V4.1 Flash 1017 waits 1.3 min for the first token
No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX

Memory for models

508.5 GB of 512 GB

120B MoE70.8 GB822B461.5 GB

For models 508.5 GBmacOS keeps 3.5 GBModel size at Q4

Everything on the size ladder fits. MiMo V2.6 Flash RL and DeepSeek V4 Flash 0731 together use 369.5 GB and leave 139 GB free.

Speed per dollar and per watt

NewUsed

On Qwen3.6 35B A3B (35B params, 3B active) at Q4_K_M. New means a shop price, either a street quote or the maker's own price; used is compared with used. The last two columns are the speed you get per $1,000 of price and per 100 W of power; higher is better. The Mac Studio (M3 Ultra, 512GB) has no current new price, so it is shown at its MSRP.

Machine	New price	tok/s	tok/s per $1,000	tok/s per 100 W
RTX 509032 GB	about $5,376	213	39.7	37.1
Mac Studio M3 Ultra96 GB	about $3,999	76.6	19.2	28.4
Mac Studio M4 Max128 GB	about $3,499	64.2	18.3	44.3
Mac Studio M2 Ultra192 GB	about $5,599	75.9	13.6	25.7
Mac Studio M3 Ultra256 GB	about $7,099	76.6	10.8	28.4
Mac Studio M3 Ultra512 GB	about $9,499	76.6	8.1	28.4

Machine	Used price	tok/s	tok/s per $1,000	tok/s per 100 W
RTX 409024 GBBest used value	about $3,000	151	50.3	33.6
RTX 509032 GB	about $6,400	213	33.3	37.1
Mac Studio M3 Ultra96 GB	about $4,525	76.6	16.9	28.4

Cost per token: the Mac Studio (M3 Ultra, 512GB) against 24–512 GB machines and GPUs

On Qwen3.6 35B A3B (35B params, 3B active, Q4_K_M, 4K context). Speed, cost per token and power per token depend on the model.

24–36 GB96–128 GB192+ GBthis desktopat MSRP

Cost per 1M output tokenson Qwen3.6 35B A3B · card over 3 years + power

$8.01

Machine, MSRP: $7.86
Power: $0.15

3-year cost of ownership$9,676

Include card price

Power$/kWhUseh/dayKeepyears

Each dot is a machine or GPU; hover or tap it for its numbers, click to open it. Machines without a price for the chosen basis are left out. How we calculate cost per token

Similar devices and alternatives

Other memory sizes of the same machine first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.6 35B A3B.

Same chip, 256 GB less memory for $2,400 lessMac Studio M3 Ultra (256 GB)Apple · 252.5 GB for models · M3 Ultra

same speed▼ $2,400 less▼ 256 GB$7,099

Same family Same chip, 416 GB less memory for $5,500 lessMac Studio M3 Ultra (96 GB)Apple · 92.5 GB for models · M3 Ultra

same speed▼ $5,500 less▼ 416 GB$3,999

Same family About as fastMac Studio M2 Ultra (192 GB)Apple · 188.5 GB for models · M2 Ultra

same speed▼ 320 GB$5,599

Step down to 124.5 GBMac Studio M4 Max (128 GB)Apple · 124.5 GB for models · M4 Max

▼ 16% speed▼ 384 GB$3,499

2.0× faster if your models fit in 24 GBRTX 4090NVIDIA · 24 GB VRAM · GPU

▲ 97% speed▼ 484.5 GB$1,599rent $0.34/hr

2.8× faster if your models fit in 32 GBRTX 5090NVIDIA · 32 GB VRAM · GPU

▲ 179% speed▼ 476.5 GB$5,376rent $0.48/hr

Mac Studio (M3 Ultra, 512GB) tokens per second by model size

Output speed depends on memory bandwidth (819 GB/s). Grey rows need more than the 508.5 GB available.

9BQwen3.5 9B · Q4_K_M · 6.6 GB
90.1tok/s
14BQwen3 14B · Q4_K_M · 11.9 GB
60.3tok/s
27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
32.9tok/s
35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
76.6tok/s
70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB
12.3tok/s
120B MoEGPT OSS 120B · Q4_K_M · 70.8 GB
75.5tok/s

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

Ultra generations: writing vs compute

M1 Ultra → M3 Ultra

Writing speed (tok/s)Compute (TFLOPS FP16)

Writing speed follows memory bandwidth and barely moved: 75.9 to 76.6 tok/s since the M1 Ultra. Compute, which sets how fast prompts are read, grew 2.1×. Indexed to the M1 Ultra; speeds on Qwen3.6 35B A3B, same 128 GB.

All 1352 models on the Mac Studio (M3 Ultra, 512GB)

Featured only

Best quant that fits in 508.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with Mac Studio (M3 Ultra, 512GB) — ranked by Score, recency and Fit
Model	Quant	VRAM	Speed	Context	Fit
Tiny MistralForCausalLM 0.1513960Chat Q4_K_M·1911.0 t/s	Q4_K_M	0.5 GB0%	1911.0 t/s	33K	S100, Runs great
Qwen3 4B Base Dapo V44.0BChat Q4_K_M·197.7 t/s	Q4_K_M	4.2 GB1%	197.7 t/s	33K	S100, Runs great
Qwen2.5 VL 7B Instruct8.3BVision Q4_K_M·106.4 t/s	Q4_K_M	6.2 GB1%	106.4 t/s	128K	S92, Runs great
Huihui Qwen3 Coder 30B A3B Instruct Abliterated30.5BChatCodeFunctions Q4_K_M·75.8 t/s	Q4_K_M	19.4 GB4%	75.8 t/s	262K	S86, Runs great
Sarashina2.2 3B Instruct v0.13.4BChat Q4_K_M·216.3 t/s	Q4_K_M	3.7 GB1%	216.3 t/s	8K	S100, Runs great
Falcon H1 0.5B Instruct521MChat Q4_K_M·830.9 t/s	Q4_K_M	1.2 GB0%	830.9 t/s	16K	S100, Runs great
SmolLM3 3B ONNX3BChat Q4_K_M·254.8 t/s	Q4_K_M	3.3 GB1%	254.8 t/s	66K	S100, Runs great
Qwen3 Embedding 0.6B596MChatFunctions Q4_K_M·744.5 t/s	Q4_K_M	1.6 GB0%	744.5 t/s	33K	S100, Runs great
Bielik 11B V3.0 Instruct11.2BChat BF16·23.3 t/s	BF16	27.4 GB5%	23.3 t/s	—	B64, Decent
Tiny FalconMambaForCausalLM1MChat BF16·57330.0 t/s	BF16	0.2 GB0%	57330.0 t/s	—	S100, Runs great
Browsesafe30.5BChat Q4_K_M·75.8 t/s	Q4_K_M	19.4 GB4%	75.8 t/s	262K	S86, Runs great
GhostFace 24B V123.6BChat Q4_K_M·38.6 t/s	Q4_K_M	17.8 GB4%	38.6 t/s	131K	A75, Runs well
Nemotron Terminal 32B32.8BChat Q4_K_M·28.3 t/s	Q4_K_M	22.6 GB4%	28.3 t/s	41K	A69, Runs well
Tiny GemmaForCausalLM2MChat Q4_K_M·1911.0 t/s	Q4_K_M	0.4 GB0%	1911.0 t/s	8K	S100, Runs great
Llm Jp 3.1 13B Instruct413.7BChat Q4_K_M·56.2 t/s	Q4_K_M	11.9 GB2%	56.2 t/s	4K	A82, Runs well

Mac Studio (M3 Ultra, 512GB) specs and power

Specs

Chip: M3 Ultra
CPU cores: 32
GPU cores: 80
Unified memory: 512 GB LPDDR5
For models: 508.5 GB (macOS keeps 3.5 GB)
Memory bandwidth: 819 GB/s
FP16 (dense): 65.5 TFLOPS
Neural Engine: 36 TOPS
NPU: 32-core Neural Engine
Power: 270 W max
Size: Mac Studio
Released: 2025-03-12
MSRP: $9,499

Mac Studio (M3 Ultra, 512GB) FAQ

Is the Mac Studio (M3 Ultra, 512GB) worth it for local LLMs in 2026?

At its $9,499 MSRP it gives 8.1 tok/s per $1,000 on Qwen3.6 35B A3B, 6th of the 6 machines compared here. It runs 27–31B and 120B MoE models well; dense 70B fits but is slow at 12 tok/s.

How much memory is available for AI on the Mac Studio (M3 Ultra, 512GB)?

About 508.5 GB of the 512 GB. macOS keeps roughly 3.5 GB for itself and other apps.

What is the largest model the Mac Studio (M3 Ultra, 512GB) can run?

Muse Spark Open 816B A42B (822.0B) is the largest at 4-bit or better: at Q4_0 it uses 461.5 GB, 91% of the available memory, and runs at about 20.0 tok/s. At 3-bit, a bigger model fits too: Inkling (952.4B) at Q3_K_M, 467.9 GB.

Can the Mac Studio (M3 Ultra, 512GB) run DeepSeek V4.1 Flash 1017?

Yes. DeepSeek V4.1 Flash 1017 fits at Q4_K_M, using 458.4 GB of the 508.5 GB available, and it is the best model that runs well on this machine.

How long does the Mac Studio (M3 Ultra, 512GB) take to read a long prompt?

On DeepSeek V4.1 Flash 1017, about 10 s for a 4K-token prompt and 80 s for a 32K-token document before the first word appears.

48 GB GPUs All Macs Desktops & mini PCs Best coding LLMs All hardware