← All Mac Studio models

Mac Studio M2 Ultra

The M2 Ultra Mac Studio runs local models at 800 GB/s of memory bandwidth with 64 to 192 GB of unified memory. On Apple Silicon that memory is shared with the GPU, so the whole pool is available for weights: at 192 GB you can hold roughly a 260B dense model at Q4. Two Max dies fused together: the widest memory bus and the highest capacity Apple sells. This is the tier that runs frontier-size open weights locally.

Apple no longer sells this configuration new. It stays fully evaluated here because the used market is where most of its local AI value now sits.

Specifications

ChipApple M2 Ultra
CPU cores24
GPU cores60 or 76
Unified memory64, 128, or 192 GB
Memory bandwidth800 GB/s
Neural Engine31.6 TOPS
Released2023
AvailabilityUsed market

Memory bandwidth is faster than 83% of the Apple Silicon chips shipped in a Mac, against a 819 GB/s peak.

Its memory ceiling is above 89% of them, against a 512 GB peak.

Pick your configuration

Every option Apple sells with this chip. The model list below recomputes against the one you pick.

GPU cores

Unified memory

What each memory option runs

Unified memory is the ceiling and it is soldered, so this is the decision you cannot revisit.

64 GB unified memory

56 GB usable for weights · 800 GB/s

3,541 of 3,641 models fit, and 3,385 of them run with headroom rather than as a squeeze.

Largest model at Q4
Kimi K2.6 JANGTQ_3L · 83.44B
Best all-round pick
Qwen3.5 4B · Q8_0 · ~90 tok/s

128 GB unified memory

112 GB usable for weights · 800 GB/s

3,582 of 3,641 models fit, and 3,532 of them run with headroom rather than as a squeeze.

Largest model at Q4
Kimi K2.5 · 171B
Best all-round pick
Qwen3.6 35B A3B · Q8_0 · ~140 tok/s

192 GB unified memory

168 GB usable for weights · 800 GB/s

3,594 of 3,641 models fit, and 3,558 of them run with headroom rather than as a squeeze.

Largest model at Q4
Llama 3_1 Nemotron Ultra 253B CPT v1 · 253.4B
Best all-round pick
Qwen3.6 35B A3B · Q8_0 · ~140 tok/s

What a 64 GB M2 Ultra Mac Studio can run

Every model in the database against this exact configuration, at 800 GB/s. Ratings and speeds are the same numbers the model pages show.

Showing 3641 of 3641 models

Qwen3.5 4B

Multimodal · Alibaba · 2026-02-27

Q8_0Excellent
5.7 GB9% of RAM~90 tok/sEstimated4.66B params
Run with ToolPiper
gemma 4 12b it

Multimodal · google · 2026-05

Q8_0Excellent
13.8 GB22% of RAM~35 tok/sEstimated11.96B params
Run with ToolPiper
Qwen3.5 0.8B

Multimodal · Alibaba · 2026-02-28

Q8_0Excellent
1.5 GB2% of RAM~482 tok/sEstimated0.87B params
Run with ToolPiper
Qwen3.5 2B

Multimodal · Alibaba · 2026-02-28

Q8_0Excellent
3.0 GB5% of RAM~185 tok/sEstimated2.27B params
Run with ToolPiper
Qwen3.5 0.8B Base

Multimodal · Alibaba · 2026-02-28

Q8_0Excellent
1.5 GB2% of RAM~482 tok/sEstimated0.87B params
Run with ToolPiper
Qwen3.5 4B Base

Multimodal · Alibaba · 2026-02-27

Q8_0Excellent
5.7 GB9% of RAM~90 tok/sEstimated4.66B params
Run with ToolPiper
Qwen3.5 2B Base

Multimodal · Alibaba · 2026-02-28

Q8_0Excellent
3.0 GB5% of RAM~185 tok/sEstimated2.27B params
Run with ToolPiper
Qwen3.5 9B Claude 4.6 Opus Reasoning Distilled v2

Reasoning · jackrong · 2026-03-16

Q8_0Excellent
11.3 GB18% of RAM~43 tok/sEstimated9.65B params
Run with ToolPiper
Qwen3.5 9B

Multimodal · Alibaba · 2026-02-27

Q8_0Excellent
11.3 GB18% of RAM~43 tok/sEstimated9.65B params
Run with ToolPiper
Qwen3.5 9B Base

Multimodal · Alibaba · 2026-02-26

Q8_0Excellent
11.3 GB18% of RAM~43 tok/sEstimated9.65B params
Run with ToolPiper
Qwen3.6 35B A3B

Multimodal · Alibaba · 2026-04-15

Q8_0Excellent
40.6 GB63% of RAMBenchmark needed35.95B params
Run with ToolPiper
LFM2 1.2B

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 ColBERT 350M

General · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LFM2.5 1.2B Thinking

Reasoning · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 8B A1B

General · Liquid AI · 2025-11-28

Q8_0Excellent
9.8 GB15% of RAMBenchmark needed8.3B params
Run with ToolPiper
LFM2 24B A2B

General · Liquid AI · 2025-11-28

Q8_0Excellent
27.1 GB42% of RAMBenchmark needed23.84B params
Run with ToolPiper
LFM2.5 1.2B Base

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 350M

General · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LFM2 2.6B

General · Liquid AI · 2025-11-28

Q8_0Excellent
3.4 GB5% of RAM~163 tok/sEstimated2.57B params
Run with ToolPiper
LFM2 700M

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.3 GB2% of RAM~566 tok/sEstimated0.74B params
Run with ToolPiper
LFM2 1.2B Extract

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 350M Extract

General · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LFM2 2.6B Transcript

General · Liquid AI · 2025-11-28

Q8_0Excellent
3.4 GB5% of RAM~163 tok/sEstimated2.57B params
Run with ToolPiper
LFM2 350M PII Extract JP

General · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LFM2 1.2B Tool

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 350M ENJP MT

General · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LFM2 1.2B RAG

General · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 350M Math

Reasoning · Liquid AI · 2025-11-28

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
LaSER Qwen3 8B

General · alibaba-nlp · 2026-03-31

Q8_0Excellent
9.6 GB15% of RAM~51 tok/sEstimated8.19B params
Run with ToolPiper
Qwen3.5 35B A3B

Multimodal · Alibaba · 2026-02-24

Q8_0Excellent
40.6 GB63% of RAMBenchmark needed35.95B params
Run with ToolPiper
granite 4.0 h tiny

General · ibm-granite · 2025-09-16

Q8_0Excellent
8.2 GB13% of RAMBenchmark needed6.94B params
Run with ToolPiper
LFM2.5 1.2B Instruct

Chat · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
LFM2 2.6B Exp

Chat · Liquid AI · 2025-11-28

Q8_0Excellent
3.4 GB5% of RAM~163 tok/sEstimated2.57B params
Run with ToolPiper
granite 4.0 h micro

General · ibm-granite · 2025-09-16

Q8_0Excellent
4.1 GB6% of RAM~131 tok/sEstimated3.19B params
Run with ToolPiper
LFM2.5 1.2B JP

Chat · Liquid AI · 2025-11-28

Q8_0Excellent
1.8 GB3% of RAM~358 tok/sEstimated1.17B params
Run with ToolPiper
Qwen3.5 35B A3B Claude 4.6 Opus Reasoning Distilled

Reasoning · jackrong · 2026-03-07

Q8_0Excellent
40.6 GB63% of RAMBenchmark needed35.95B params
Run with ToolPiper
VAETKI 7B A1B

General · NCAI · 2025-12-29

Q8_0Excellent
8.6 GB13% of RAMBenchmark needed7.25B params
Run with ToolPiper
VAETKI 20B A2B

General · NCAI · 2025-12-29

Q8_0Excellent
22.4 GB35% of RAMBenchmark needed19.6B params
Run with ToolPiper
gemma 4 26B A4B it

Multimodal · Google · 2025-07-30

Q8_0Excellent
29.5 GB46% of RAMBenchmark needed26B params
Run with ToolPiper
gemma 4 E4B it

Multimodal · Google · 2025-07-30

Q8_0Excellent
9.4 GB15% of RAM~52 tok/sEstimated8B params
Run with ToolPiper
gemma 4 E2B it

Multimodal · Google · 2025-07-30

Q8_0Excellent
6.2 GB10% of RAM~82 tok/sEstimated5.1B params
Run with ToolPiper
SmolLM3 3B

Reasoning · HuggingFace · 2025-07-08

Q8_0Excellent
3.8 GB6% of RAM~140 tok/sEstimated3B params
Run with ToolPiper
granite 4.0 h small

General · ibm-granite · 2025-09-16

Q8_0Excellent
36.4 GB57% of RAMBenchmark needed32.21B params
Run with ToolPiper
gemma 3n E2B it

Multimodal · Google · 2025-06-25

Q8_0Excellent
5.0 GB8% of RAM~105 tok/sEstimated4B params
Run with ToolPiper
LFM2.5 Audio 1.5B

General · Liquid AI · 2025-11-28

Q8_0Excellent
2.2 GB3% of RAM~279 tok/sEstimated1.5B params
Run with ToolPiper
LFM2 Audio 1.5B

General · Liquid AI · 2025-11-28

Q8_0Excellent
2.2 GB3% of RAM~279 tok/sEstimated1.5B params
Run with ToolPiper
VAETKI VL 7B A1B

Multimodal · NCAI · 2025-12-29

Q8_0Excellent
9.0 GB14% of RAMBenchmark needed7.58B params
Run with ToolPiper
NVIDIA Nemotron Nano 9B v2

Reasoning · NVIDIA · 2025-06-01

Q8_0Excellent
10.5 GB16% of RAM~47 tok/sEstimated9B params
Run with ToolPiper
LFM2 VL 450M

Multimodal · Liquid AI · 2025-11-28

Q8_0Excellent
1.0 GB2% of RAM~931 tok/sEstimated0.45B params
Run with ToolPiper
LFM2.5 VL 1.6B

Multimodal · Liquid AI · 2025-11-28

Q8_0Excellent
2.3 GB4% of RAM~262 tok/sEstimated1.6B params
Run with ToolPiper
LFM2 VL 3B

Multimodal · Liquid AI · 2025-11-28

Q8_0Excellent
3.8 GB6% of RAM~140 tok/sEstimated3B params
Run with ToolPiper
LFM2 VL 1.6B

Multimodal · Liquid AI · 2025-11-28

Q8_0Excellent
2.3 GB4% of RAM~265 tok/sEstimated1.58B params
Run with ToolPiper
EXAONE Deep 2.4B

General · lgai-exaone · 2025-03-12

Q8_0Excellent
3.2 GB5% of RAM~174 tok/sEstimated2.41B params
Run with ToolPiper
DeepSeek R1 Distill Qwen 7B

Reasoning · DeepSeek · 2025-01-20

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled

Reasoning · jackrong · 2026-02-27

Q8_0Excellent
31.5 GB49% of RAM~15 tok/sEstimated27.78B params
Run with ToolPiper
EXAONE 4.0 1.2B

General · LG AI · 2025-07-15

Q8_0Excellent
1.8 GB3% of RAM~349 tok/sEstimated1.2B params
Run with ToolPiper
Jolia

General · raidium · 2026-06-15

Q8_0Excellent
0.5 GB1% of RAM~20,952 tok/sEstimated0.02B params
Run with ToolPiper
embeddinggemma GTAIDE 300m 2605

Embedding · taide · 2026-06-12

Q8_0Excellent
0.8 GB1% of RAM~1,397 tok/sEstimated0.3B params
Run with ToolPiper
Qwen3 0.6B

General · Alibaba · 2025-04-27

Q8_0Excellent
1.3 GB2% of RAM~559 tok/sEstimated0.75B params
Run with ToolPiper
Qwen3 4B

General · Alibaba · 2025-04-27

Q8_0Excellent
5.0 GB8% of RAM~104 tok/sEstimated4.02B params
Run with ToolPiper
gpt oss 20b

General · openai

Q8_0Excellent
24.5 GB38% of RAMBenchmark needed21.51B params
Run with ToolPiper
Qwen3.6 27B

Multimodal · Alibaba · 2026-04-21

Q8_0Excellent
31.5 GB49% of RAM~15 tok/sEstimated27.78B params
Run with ToolPiper
Qwen3 1.7B

General · Alibaba · 2025-04-27

Q8_0Excellent
2.8 GB4% of RAM~206 tok/sEstimated2.03B params
Run with ToolPiper
Qwen3 VL 4B Instruct

Multimodal · Alibaba

Q8_0Excellent
5.5 GB9% of RAM~94 tok/sEstimated4.44B params
Run with ToolPiper
GLM OCR

Multimodal · zai-org

Q8_0Excellent
2.0 GB3% of RAM~315 tok/sEstimated1.33B params
Run with ToolPiper
Qwen3 30B A3B

General · Alibaba · 2025-04-27

Q8_0Excellent
34.6 GB54% of RAMBenchmark needed30.53B params
Run with ToolPiper
gemma 3 12b it

Multimodal · Google · 2025-03-01

Q8_0Excellent
13.9 GB22% of RAM~35 tok/sEstimated12B params
Run with ToolPiper
Qwen3 VL 2B Instruct

Multimodal · Alibaba

Q8_0Excellent
2.9 GB4% of RAM~197 tok/sEstimated2.13B params
Run with ToolPiper
Qwen3 Coder 30B A3B Instruct

Coding · Alibaba

Q8_0Excellent
34.6 GB54% of RAMBenchmark needed30.53B params
Run with ToolPiper
chandra ocr 2

Multimodal · datalab-to

Q8_0Excellent
6.4 GB10% of RAM~79 tok/sEstimated5.3B params
Run with ToolPiper
Rio 3.0 Open Mini

General · prefeitura-rio

Q8_0Excellent
5.0 GB8% of RAM~104 tok/sEstimated4.02B params
Run with ToolPiper
Phi 3.5 vision instruct

Multimodal · Microsoft

Q8_0Excellent
5.1 GB8% of RAM~101 tok/sEstimated4.15B params
Run with ToolPiper
Qwen3 VL 4B Thinking

General · Alibaba

Q8_0Excellent
5.5 GB9% of RAM~94 tok/sEstimated4.44B params
Run with ToolPiper
Qwen2.5 1.5B

General · Alibaba

Q8_0Excellent
2.2 GB3% of RAM~272 tok/sEstimated1.54B params
Run with ToolPiper
Qwen2 0.5B

General · Alibaba

Q8_0Excellent
1.0 GB2% of RAM~855 tok/sEstimated0.49B params
Run with ToolPiper
diffusiongemma 26B A4B it

Multimodal · Google

Q8_0Excellent
29.3 GB46% of RAMBenchmark needed25.82B params
Run with ToolPiper
MiniCPM V 4.6

Multimodal · openbmb

Q8_0Excellent
2.0 GB3% of RAM~322 tok/sEstimated1.3B params
Run with ToolPiper
Qwen3 4B Thinking 2507

General · Alibaba

Q8_0Excellent
5.0 GB8% of RAM~104 tok/sEstimated4.02B params
Run with ToolPiper
DeepSeek R1 Distill Qwen 1.5B

Reasoning · DeepSeek

Q8_0Excellent
2.5 GB4% of RAM~235 tok/sEstimated1.78B params
Run with ToolPiper
dots.mocr

Multimodal · rednote-hilab

Q8_0Excellent
3.9 GB6% of RAM~138 tok/sEstimated3.04B params
Run with ToolPiper
Qwen2.5 7B

General · Alibaba

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
Phi 4 multimodal instruct

Multimodal · Microsoft · 2025-04-01

Q8_0Excellent
16.1 GB25% of RAM~30 tok/sEstimated14B params
Run with ToolPiper
DeepSeek R1 0528 Qwen3 8B

Reasoning · DeepSeek

Q8_0Excellent
9.6 GB15% of RAM~51 tok/sEstimated8.19B params
Run with ToolPiper
UI TARS 1.5 7B

Multimodal · bytedance-seed

Q8_0Excellent
9.7 GB15% of RAM~51 tok/sEstimated8.29B params
Run with ToolPiper
VLM2Vec Full

General · tiger-lab

Q8_0Excellent
5.1 GB8% of RAM~101 tok/sEstimated4.15B params
Run with ToolPiper
DeepSeek V2 Lite

General · DeepSeek

Q8_0Excellent
18.0 GB28% of RAMBenchmark needed15.71B params
Run with ToolPiper
gemma 4 26B A4B

Multimodal · Google

Q8_0Excellent
30.1 GB47% of RAMBenchmark needed26.54B params
Run with ToolPiper
surya ocr 2

Multimodal · datalab-to

Q8_0Excellent
1.3 GB2% of RAM~607 tok/sEstimated0.69B params
Run with ToolPiper
granite 4.1 3b

General · ibm-granite

Q8_0Excellent
4.3 GB7% of RAM~123 tok/sEstimated3.4B params
Run with ToolPiper
DeepSeek R1 Distill Llama 8B

Reasoning · DeepSeek

Q8_0Excellent
9.5 GB15% of RAM~52 tok/sEstimated8.03B params
Run with ToolPiper
OTel LLM 4B IT

General · farbodtavakkoli

Q8_0Excellent
5.3 GB8% of RAM~97 tok/sEstimated4.3B params
Run with ToolPiper
Kimi VL A3B Instruct

Multimodal · moonshotai

Q8_0Excellent
18.8 GB29% of RAMBenchmark needed16.41B params
Run with ToolPiper
MiniCPM5 1B

General · openbmb

Q8_0Excellent
1.7 GB3% of RAM~388 tok/sEstimated1.08B params
Run with ToolPiper
OTel LLM 8.3B IT

General · farbodtavakkoli

Q8_0Excellent
8.3 GB13% of RAM~60 tok/sEstimated6.97B params
Run with ToolPiper
dots.ocr

Multimodal · rednote-hilab

Q8_0Excellent
3.9 GB6% of RAM~138 tok/sEstimated3.04B params
Run with ToolPiper
Ilama 3.2 1B

General · hmellor

Q8_0Excellent
1.9 GB3% of RAM~338 tok/sEstimated1.24B params
Run with ToolPiper
distil lfm25 shellper

General · distil-labs

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
Nanonets OCR2 3B

Multimodal · nanonets

Q8_0Excellent
4.7 GB7% of RAM~112 tok/sEstimated3.75B params
Run with ToolPiper
gemma 4 26B A4B it uncensored

General · trevorjs

Q8_0Excellent
29.3 GB46% of RAMBenchmark needed25.81B params
Run with ToolPiper
OTel LLM 1.2B IT

General · farbodtavakkoli

Q8_0Excellent
1.8 GB3% of RAM~346 tok/sEstimated1.21B params
Run with ToolPiper
RolmOCR

Multimodal · reducto

Q8_0Excellent
9.7 GB15% of RAM~51 tok/sEstimated8.29B params
Run with ToolPiper
Qwen3 30B A3B Thinking 2507

General · Alibaba

Q8_0Excellent
34.6 GB54% of RAMBenchmark needed30.53B params
Run with ToolPiper
LFM2.5 8B A1B

General · Liquid AI

Q8_0Excellent
9.9 GB16% of RAMBenchmark needed8.47B params
Run with ToolPiper
Nanonets OCR s

General · nanonets

Q8_0Excellent
4.7 GB7% of RAM~112 tok/sEstimated3.75B params
Run with ToolPiper
Nanbeige4.1 3B

General · nanbeige

Q8_0Excellent
4.9 GB8% of RAM~107 tok/sEstimated3.93B params
Run with ToolPiper
LLaVA OneVision 1.5 4B Base

General · lmms-lab

Q8_0Excellent
5.8 GB9% of RAM~88 tok/sEstimated4.74B params
Run with ToolPiper
granite vision 4.1 4b

Multimodal · ibm-granite

Q8_0Excellent
5.0 GB8% of RAM~105 tok/sEstimated4B params
Run with ToolPiper
Molmo2 O 7B

Multimodal · allenai

Q8_0Excellent
9.2 GB14% of RAM~54 tok/sEstimated7.76B params
Run with ToolPiper
OTel LLM 3B IT

General · farbodtavakkoli

Q8_0Excellent
5.2 GB8% of RAM~99 tok/sEstimated4.25B params
Run with ToolPiper
granite 4.0 3b vision

General · ibm-granite

Q8_0Excellent
5.0 GB8% of RAM~105 tok/sEstimated4B params
Run with ToolPiper
gpt oss 20b Derestricted

General · arliai

Q8_0Excellent
23.8 GB37% of RAMBenchmark needed20.91B params
Run with ToolPiper
granite 4.0 tiny preview

General · ibm-granite

Q8_0Excellent
7.9 GB12% of RAMBenchmark needed6.67B params
Run with ToolPiper
reformer crime and punishment

General · Google

Q8_0Excellent
0.5 GB1% of RAM~41,905 tok/sEstimated0.01B params
Run with ToolPiper
Kimi VL A3B Thinking

Multimodal · moonshotai

Q8_0Excellent
18.8 GB29% of RAMBenchmark needed16.41B params
Run with ToolPiper
VideoLLaMA3 2B Image HF

Multimodal · lkhl

Q8_0Excellent
2.7 GB4% of RAM~214 tok/sEstimated1.96B params
Run with ToolPiper
typhoon ocr 3b

Multimodal · typhoon-ai

Q8_0Excellent
4.7 GB7% of RAM~112 tok/sEstimated3.75B params
Run with ToolPiper
ReaderLM v2

General · jinaai

Q8_0Excellent
2.2 GB3% of RAM~272 tok/sEstimated1.54B params
Run with ToolPiper
Qwen3 VL 2B Thinking

General · Alibaba

Q8_0Excellent
2.9 GB4% of RAM~197 tok/sEstimated2.13B params
Run with ToolPiper
Josiefied Qwen3 VL 4B Instruct abliterated beta v1

Multimodal · goekdeniz-guelmez

Q8_0Excellent
5.5 GB9% of RAM~94 tok/sEstimated4.44B params
Run with ToolPiper
gemma 4 26B A4B it heretic

Coding · coder3101

Q8_0Excellent
29.3 GB46% of RAMBenchmark needed25.81B params
Run with ToolPiper
dots.ocr 1.5

General · kristaller486

Q8_0Excellent
3.9 GB6% of RAM~138 tok/sEstimated3.04B params
Run with ToolPiper
typhoon ocr1.5 2b

Reasoning · typhoon-ai

Q8_0Excellent
2.9 GB4% of RAM~197 tok/sEstimated2.13B params
Run with ToolPiper
Solar Open 100B

General · Upstage

Q8_0Excellent
9.5 GB15% of RAMBenchmark needed8.05B params
Run with ToolPiper
Yi 6B 200K

General · 01.ai

Q8_0Excellent
7.3 GB11% of RAM~69 tok/sEstimated6.06B params
Run with ToolPiper
HTML Pruner Phi 3.8B

General · zstanjj

Q8_0Excellent
4.8 GB7% of RAM~110 tok/sEstimated3.82B params
Run with ToolPiper
granite 4.1 3b base

General · ibm-granite

Q8_0Excellent
4.3 GB7% of RAM~123 tok/sEstimated3.4B params
Run with ToolPiper
Nemotron Labs Diffusion 3B

General · nvidia

Q8_0Excellent
4.8 GB7% of RAM~109 tok/sEstimated3.83B params
Run with ToolPiper
VibeThinker 3B

General · weiboai

Q8_0Excellent
3.9 GB6% of RAM~136 tok/sEstimated3.09B params
Run with ToolPiper
granite 8b code instruct 128k

Coding · ibm-granite

Q8_0Excellent
9.5 GB15% of RAM~52 tok/sEstimated8.05B params
Run with ToolPiper
Qwen2 7B

General · Alibaba

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
libra v1.0 3b

General · x-izhang

Q8_0Excellent
4.1 GB6% of RAM~129 tok/sEstimated3.25B params
Run with ToolPiper
Rex Omni

General · idea-research

Q8_0Excellent
5.0 GB8% of RAM~103 tok/sEstimated4.07B params
Run with ToolPiper
gpt oss safeguard 120b

General · openai

Q8_0Excellent
3.1 GB5% of RAMBenchmark needed2.37B params
Run with ToolPiper
PaddleOCR VL 1.5

General · paddlepaddle

Q8_0Excellent
1.6 GB2% of RAM~437 tok/sEstimated0.96B params
Run with ToolPiper
ue3 qw35 4b v1

General · tristepin

Q8_0Excellent
5.7 GB9% of RAM~90 tok/sEstimated4.66B params
Run with ToolPiper
gemma 3 4B T1 it

General · twinkle-ai

Q8_0Excellent
4.8 GB8% of RAM~108 tok/sEstimated3.88B params
Run with ToolPiper
LFM2.5 350M Base

General · Liquid AI

Q8_0Excellent
0.9 GB1% of RAM~1,197 tok/sEstimated0.35B params
Run with ToolPiper
plamo 2 1b

General · pfnet

Q8_0Excellent
1.9 GB3% of RAM~325 tok/sEstimated1.29B params
Run with ToolPiper
Dream v0 Base 7B

General · dream-org

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
Foundation Sec 8B Reasoning

Reasoning · fdtn-ai

Q8_0Excellent
9.5 GB15% of RAM~52 tok/sEstimated8.03B params
Run with ToolPiper
AceReason Nemotron 7B

Reasoning · nvidia

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
AceReason Nemotron 1.1 7B

Reasoning · nvidia

Q8_0Excellent
9.0 GB14% of RAM~55 tok/sEstimated7.62B params
Run with ToolPiper
AFM 4.5B Base

General · arcee-ai

Q8_0Excellent
5.7 GB9% of RAM~91 tok/sEstimated4.62B params
Run with ToolPiper
MiniCPM5 1B SFT

General · openbmb

Q8_0Excellent
1.7 GB3% of RAM~388 tok/sEstimated1.08B params
Run with ToolPiper
granite 4.0 micro base

General · ibm-granite

Q8_0Excellent
4.3 GB7% of RAM~123 tok/sEstimated3.4B params
Run with ToolPiper
ERNIE 4.5 0.3B PT

General · baidu

Q8_0Excellent
0.9 GB1% of RAM~1,164 tok/sEstimated0.36B params
Run with ToolPiper
Mellum2 12B A2.5B Thinking

General · jetbrains

Q8_0Excellent
14.1 GB22% of RAMBenchmark needed12.15B params
Run with ToolPiper
PARD Llama 3.2 1B

General · amd

Q8_0Excellent
2.2 GB3% of RAM~279 tok/sEstimated1.5B params
Run with ToolPiper
academic ds 9B

General · bytedance-seed

Q8_0Excellent
11.0 GB17% of RAMBenchmark needed9.37B params
Run with ToolPiper
Jan nano 128k

General · menlo

Q8_0Excellent
5.0 GB8% of RAM~104 tok/sEstimated4.02B params
Run with ToolPiper

60-core vs 76-core GPU

Both bins run the same 800 GB/s memory bus, so token generation is the same on either one. The extra cores show up in image and video work, not in tokens per second.

ConfigurationMemory bandwidthMemory optionsModels that fit
24-core CPU, 60-core GPU800 GB/s64, 128, 192 GBIdentical
24-core CPU, 76-core GPU800 GB/s64, 128, 192 GBIdentical

Measured on the M2 Ultra

Nobody has submitted a benchmark on the M2 Ultra yet, so every speed on this page is the formula estimate rather than a measured run. The estimate is bandwidth-driven and calibrated against chips that do have data, which makes it a good guide and not a promise.

ToolPiper contributes a result anonymously when you run the benchmark, and the leaderboard shows every chip that already has one.

Where to go from here

What each step actually changes for local models, rather than which one is newer.

Newer generation

Mac Studio M3 Ultra

1x the memory bandwidth, up to 512 GB instead of 192 GB

Used market alternative

Mac Studio M1 Ultra

128 GB ceiling instead of 192 GB

Why unified memory is the number that matters

On a PC the model has to fit in GPU VRAM, which is a separate pool from system RAM and usually the smaller of the two. Apple Silicon has one pool. The M2 Ultra's 800 GB/s bus is shared by CPU, GPU, and Neural Engine, so a 192 GB machine can hand almost all of that to a model with no copy across a bus.

Apple stopped selling this one, which is exactly why it is interesting. The Studio exists for this workload. It carries the widest memory buses and the highest capacities Apple sells, and it runs at full clocks indefinitely. A used M2 Ultra at 192 GB still gives you 800 GB/s and a hard 260B ceiling, and neither number degrades with age the way a battery does.

Common questions

Can the M2 Ultra Mac Studio run a 70B model?

Yes, at 192 GB. A 70B model at Q4_K_M needs about 46 GB including an 8K context, and 192 GB of unified memory leaves about 168 GB for weights once macOS takes its share. At 64 GB it does not fit at any quantization worth running.

How much unified memory should I get with the M2 Ultra Mac Studio?

Memory is the only spec that changes what you can run at all. 64 GB holds about a 85B model at Q4; 192 GB holds about 260B. It is soldered, so this is a one-time decision, and it is the upgrade worth paying for before core count.

How fast are local LLMs on the M2 Ultra?

Token generation is bandwidth-bound, so M2 Ultra throughput scales with its 800 GB/s memory bus. Divide bandwidth by the size of the weights actually read per token to get the ceiling, then expect roughly half of that in practice. A 7B model at Q4 reads about 4 GB per token pass, so the M2 Ultra lands in the tens of tokens per second and a 70B model lands in the single digits.

Is the 76-core GPU worth it over the 60-core on the M2 Ultra?

Not for LLMs. Both bins run the same 800 GB/s memory bus and take the same memory options, and token generation is bound by bandwidth rather than GPU cores. The extra cores show up in image generation and video work, not in tokens per second.

Is a used M2 Ultra Mac Studio still worth buying for local AI?

For inference, the specs that matter do not age: 800 GB/s and up to 192 GB of unified memory are the same numbers today as they were in 2023. A used M2 Ultra at the top memory option usually beats a new base-tier machine at the same price on both. Check the battery and the display, not the silicon.

Run these models on your Mac Studio

ToolPiper downloads, manages, and runs local models on Apple Silicon. Free, and nothing leaves the machine.

Get ToolPiperCompare against another Mac