Hardware · Clusters · Example cluster

Two Mac Studios (M3 Ultra, 256 GB): what it runs together

Two 256 GB Mac Studios on Thunderbolt 5 with RDMA.

About 380 GB to work with: GLM 4.7 (358B) at 4-bit instead of 2-bit, and DeepSeek V3.1 (685B) at 3-bit, which neither Mac holds alone. With RDMA, EXO's tensor split also types dense models faster than one Mac.

Change the devices How clustering works

384 GBusable memory together (the largest device has 192 GB)

21models only the cluster can run

364run better or faster than on any one device

822B (42.0B active)biggest model it adds: Muse Spark Open (816B-A42B)

Connected by Thunderbolt 5 with RDMA (macOS 26.2+). The wait for the first word is worked out for a 1,000-token prompt (a short chat).

The devices

Tool to use

EXO best estimate for 378 of the 385 models the cluster adds

Two or more Apple silicon Macs, best of all Mac Studios or Mac minis joined by Thunderbolt 5 cables, when you want big models with almost no setup.

brew install --cask exo   # on every Mac, then open EXO
open http://localhost:52415   # pick a model and how to split it

601 LLMs this cluster can run

Models the cluster adds come first, best match first, then the ones a single device already runs as well. Filter vs one device to pick. Speeds are estimates for one person chatting.

Quality Model Quant Memory Speed Runs Devices Tool vs one device
92Excellent quality (A+) · 92 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. Ling 3.1 Flash 1019 (560B-A25B)Ant Group (inclusionAI) · 572B (25.0B active) Q4_K_M 381 GB 50.7 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
90Excellent quality (A+) · 90 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. Ling 3.1 Flash (560B-A25B)Ant Group (inclusionAI) · 572B (25.0B active) Q4_K_M 381 GB 50.7 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
83Great quality (A) · 83 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 91/100 (quality, fit, speed and memory use) sets the ranking. LongCat Flash Chat (560B)Meituan · 562B (27.0B active) Q4_K_M 375 GB 42.1 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
88Excellent quality (A+) · 88 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 91/100 (quality, fit, speed and memory use) sets the ranking. Nemotron 3 Ultra 550B-A55BNVIDIA · 561B (55.0B active) Q4_K_M 374 GB 26.6 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
81Great quality (A) · 81 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 90/100 (quality, fit, speed and memory use) sets the ranking. Step 5 Preview (600B-A27B)StepFun · 604B (27.0B active) Q3_K_M 329 GB 52.1 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
78Great quality (A) · 78 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 88/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3.2 (671B)DeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
78Great quality (A) · 78 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 88/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek R1 0528 (671B)DeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
77Great quality (A) · 77 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 87/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3.2 ExpDeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
77Great quality (A) · 77 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 87/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek R1 (671B)DeepSeek · 671B (37.0B active) Q3_K_M 361 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
76Great quality (A) · 76 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 87/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3.1 TerminusDeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
76Great quality (A) · 76 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 87/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3.1 (671B)DeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
75Great quality (A) · 75 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. Mistral Large 3 (675B)Mistral AI · 675B (41.0B active) Q3_K_M 363 GB 32.3 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
68Good quality (B) · 68 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 84/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V4.1 Flash 1017 (552B-A16B)DeepSeek · 763B (16.0B active) Q2_K 337 GB 55.7 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
69Good quality (B) · 69 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 84/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3 0324 (671B)DeepSeek · 685B (37.0B active) Q3_K_M 368 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
68Good quality (B) · 68 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking. GLM-5.3 (753B-A40B)Zhipu AI · 753B (40.0B active) Q2_K 334 GB 28.7 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
67Good quality (B) · 67 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking. GLM-5.2 (753B-A40B)Zhipu AI · 753B (40.0B active) Q2_K 334 GB 28.7 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
67Good quality (B) · 67 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. Muse Spark Open (816B-A42B)Meta · 822B (42.0B active) Q2_K 361 GB 27.6 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
66Good quality (B) · 66 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. GLM-5.1 (744B-A40B)Zhipu AI · 754B (40.0B active) Q2_K 334 GB 28.7 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
66Good quality (B) · 66 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. GLM-5 (744B-A40B)Zhipu AI · 754B (40.0B active) Q2_K 334 GB 28.7 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
63Good quality (B) · 63 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V3 (671B)DeepSeek · 671B (37.0B active) Q3_K_M 361 GB 34.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Needs the clusterWon’t run on any one of them
67Good quality (B) · 67 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking. Hunyuan Hy4 Preview (780B)Tencent · 780B (55.0B active) Q2_K 345 GB 22.0 tok/s Runs welltight fit Mac 1 + Mac 2Pipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
95Excellent quality (A+) · 95 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 98/100 (quality, fit, speed and memory use) sets the ranking. GLM-5.3 Flash (320B-A18B)Zhipu AI · 321B (18.0B active) Q6_K 295 GB 47.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 51.7 tok/s
95Excellent quality (A+) · 95 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 98/100 (quality, fit, speed and memory use) sets the ranking. MiMo V2.6 Flash (310B-A15B)Xiaomi · 311B (15.0B active) Q6_K 287 GB 54.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 60.1 tok/s
94Excellent quality (A+) · 94 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 97/100 (quality, fit, speed and memory use) sets the ranking. Naive N0.5 Flash (309B-A15.5B)Naive AI · 309B (15.5B active) Q6_K 285 GB 53.5 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 58.5 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 97/100 (quality, fit, speed and memory use) sets the ranking. Inkling Small (276B-A12B)Thinking Machines Lab · 266B (12.0B active) Q8_0 314 GB 51.6 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 58.0 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V4 Flash (284B-A13B)DeepSeek · 304B (13.0B active) Q6_K 282 GB 45.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 67.3 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. MiMo V2.5 (310B-A15B)Xiaomi · 311B (15.0B active) Q6_K 287 GB 54.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 60.1 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Step 3.7 Flash (198B-A11B)StepFun · 201B (11.0B active) Q8_0 246 GB 52.4 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 53.0 tok/s
96Excellent quality (A+) · 96 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Qwen3.8 Flash Next (125B-A6B)Alibaba · 180B (6.0B active) Q8_0 223 GB 75.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 84.1 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M2.7 (230B-A10B)MiniMax · 229B (10.0B active) Q8_0 276 GB 38.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 67.0 tok/s
94Excellent quality (A+) · 94 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Intern-S2 (397B-A17B)Shanghai AI Lab · 403B (17.0B active) Q6_K 363 GB 48.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 66.3 tok/s
94Excellent quality (A+) · 94 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Nex N2 Pro (397B-A17B)Nex-AGI · 397B (17.0B active) Q6_K 358 GB 48.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 66.3 tok/s
94Excellent quality (A+) · 94 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Qwen3.5 397B-A17BAlibaba · 403B (17.0B active) Q6_K 363 GB 48.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 66.3 tok/s
92Excellent quality (A+) · 92 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M2.5 (230B-A10B)MiniMax · 229B (10.0B active) Q8_0 276 GB 38.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 67.0 tok/s
92Excellent quality (A+) · 92 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M3.1 Flash (428B-A23B)MiniMax · 429B (23.0B active) Q4_K_M 295 GB 38.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 51.8 tok/s
92Excellent quality (A+) · 92 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 96/100 (quality, fit, speed and memory use) sets the ranking. Step 3.5 Flash (196B-A11B)StepFun · 199B (11.0B active) Q8_0 244 GB 52.4 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 53.0 tok/s
91Excellent quality (A+) · 91 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek V4 Flash (April 2026)DeepSeek · 291B (13.0B active) Q6_K 271 GB 45.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 67.3 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. Ornith 1.5 397B-A17BDeepReinforce · 403B (17.0B active) Q6_K 363 GB 48.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 66.3 tok/s
91Excellent quality (A+) · 91 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M2.1 (230B-A10B)MiniMax · 229B (10.0B active) Q8_0 276 GB 38.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 67.0 tok/s
89Excellent quality (A+) · 89 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. MiMo V2 Flash (309B-A15B)Xiaomi · 310B (15.0B active) Q6_K 286 GB 54.7 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 60.1 tok/s
89Excellent quality (A+) · 89 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 95/100 (quality, fit, speed and memory use) sets the ranking. K2-Horizon (375B-A23B)Institute of Foundation Models · 379B (23.0B active) Q4_K_M 265 GB 37.9 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 51.8 tok/s
90Excellent quality (A+) · 90 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M3 (427B)MiniMax · 427B (25.0B active) Q4_K_M 293 GB 36.5 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 48.3 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. Hunyuan Hy3 (295B-A21B)Tencent · 299B (21.0B active) Q6_K 280 GB 29.0 tok/s Runs well Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 45.4 tok/s
89Excellent quality (A+) · 89 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. MiniMax M2 (230B-A10B)MiniMax · 229B (10.0B active) Q8_0 276 GB 38.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 67.0 tok/s
89Excellent quality (A+) · 89 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. Solar Open 2 250BUpstage · 250B (15.0B active) Q6_K 239 GB 41.0 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs well on Mac 1, 48.3 tok/s
89Excellent quality (A+) · 89 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 94/100 (quality, fit, speed and memory use) sets the ranking. Beam 1.0 (501B-A23B)Reflection AI · 501B (23.4B active) Q4_K_M 338 GB 53.2 tok/s Runs welltight fit Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Slow on Mac 1, 43.4 tok/s
87Excellent quality (A+) · 87 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking. K-EXAONE 236B-A23BLG AI Research · 237B (23.0B active) Q6_K 227 GB 39.3 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 33.4 tok/s
88Excellent quality (A+) · 88 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking. GLM-4.7 (355B-A32B)Zhipu AI · 358B (32.0B active) Q4_K_M 254 GB 27.6 tok/s Runs well Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 39.0 tok/s
93Excellent quality (A+) · 93 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 92/100 (quality, fit, speed and memory use) sets the ranking. Nemotron 3.5 Super VL 120B-A12BNVIDIA · 124B (12.0B active) Q8_0 164 GB 51.6 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 42.4 tok/s
92Excellent quality (A+) · 92 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 91/100 (quality, fit, speed and memory use) sets the ranking. Ornith 1.5 122B-A10BDeepReinforce · 125B (10.0B active) Q8_0 165 GB 56.2 tok/s Runs great Mac 1 + Mac 2Tensor split EXO Better on the clusterAlone: Runs great on Mac 1, 49.4 tok/s

Other example clusters

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.

Was this page helpful?