Hardware · Clusters · Example cluster

MacBook Pro + gaming PC + Raspberry Pi: what it runs together

A MacBook Pro with 48 GB, a gaming PC with an 8 GB RTX 4060 and a Raspberry Pi 4 on the same home network.

Together the Mac, the PC's graphics card and its system RAM run 100B-class mixture-of-experts models none of them can hold alone, at about 15 to 26 tokens per second, and 70B models at a better quant than the Mac manages alone. The Raspberry Pi adds memory but reads it more than 40 times slower than the Mac, so it is needed for almost nothing: leave it out.

Change the devices How clustering works

50.4 GBusable memory together (the largest device has 36.0 GB)

28models only the cluster can run

23run better or faster than on any one device

143B (14.0B active)biggest model it adds: dots.llm1 Instruct

Connected by 1 Gb Ethernet. The wait for the first word is worked out for a 1,000-token prompt (a short chat).

The devices

Tool to use

llama.cpp RPC best estimate for 51 of the 51 models the cluster adds

Mixed devices you already own, a Mac, a gaming PC, a mini PC or a Jetson, when one GGUF model won't fit on any of them alone.

ggml-rpc-server --host 0.0.0.0 --port 50052 -c   # on each worker
llama-server -hf <model-GGUF> -ngl 99 --rpc 192.168.1.10:50052,192.168.1.11:50052   # on the main machine

Also works. The number is how many of the added models each one runs:

518 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
65Good quality (B) · 65 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking. Ring Flash 2.0 (100B-A6B)Ant Group (inclusionAI) · 103B (6.1B active) Q3_K_M 55.4 GB 26.2 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline 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 65/100 (quality, fit, speed and memory use) sets the ranking. Ling Flash 2.0 (100B-A6B)Ant Group (inclusionAI) · 103B (6.1B active) Q3_K_M 55.4 GB 26.2 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline 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 65/100 (quality, fit, speed and memory use) sets the ranking. GLM-4.6VZhipu AI · 108B (12.0B active) Q3_K_M 58.8 GB 15.0 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
64Good quality (B) · 64 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 65/100 (quality, fit, speed and memory use) sets the ranking. Ornith 1.5 122B-A10BDeepReinforce · 125B (10.0B active) Q2_K 54.7 GB 22.4 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
64Good quality (B) · 64 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 65/100 (quality, fit, speed and memory use) sets the ranking. Qwen3.5 122B-A10BAlibaba · 125B (10.0B active) Q2_K 54.7 GB 22.4 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
65Good quality (B) · 65 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 65/100 (quality, fit, speed and memory use) sets the ranking. Nemotron 3.5 Super VL 120B-A12BNVIDIA · 124B (12.0B active) Q2_K 54.3 GB 19.7 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
57Good quality (B) · 57 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking. Ling 3.0 Flash (127B-A5B)Ant Group (inclusionAI) · 127B (5.1B active) Q2_K 55.9 GB 34.2 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline 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 Q3_K_M, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking. Solar Open 100BUpstage · 103B (12.0B active) Q3_K_M 56.4 GB 15.7 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline 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 Q3_K_M, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking. GLM-4.5VZhipu AI · 108B (12.0B active) Q3_K_M 58.8 GB 15.0 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
60Good quality (B) · 60 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking. Laguna S 2.1 (118B-A8B)Poolside · 118B (8.0B active) Q2_K 53.1 GB 27.0 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
56Good quality (B) · 56 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 63/100 (quality, fit, speed and memory use) sets the ranking. Mistral Small 4 (119B-A6B)Mistral AI · 119B (6.0B active) Q2_K 52.4 GB 32.8 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
65Good quality (B) · 65 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 63/100 (quality, fit, speed and memory use) sets the ranking. GLM-4.5 Air (106B-A12B)Zhipu AI · 110B (12.0B active) Q3_K_M 60.2 GB 14.7 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline 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 63/100 (quality, fit, speed and memory use) sets the ranking. Sarvam 105BSarvam AI · 106B (10.0B active) Q3_K_M 56.7 GB 17.9 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
61Good quality (B) · 61 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 63/100 (quality, fit, speed and memory use) sets the ranking. Nemotron 3 Super 120B-A12BNVIDIA · 124B (12.0B active) Q2_K 54.0 GB 19.8 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
49Fair quality (C) · 49 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 60/100 (quality, fit, speed and memory use) sets the ranking. Groot 65B A6B 261023GrootTeam · 118B (6.0B active) Q2_K 52.0 GB 33.0 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
47Fair quality (C) · 47 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 60/100 (quality, fit, speed and memory use) sets the ranking. Ling 3.0 flash SanteAnt Group (inclusionAI) · 127B (5.1B active) Q2_K 55.5 GB 34.2 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
60Good quality (B) · 60 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 60/100 (quality, fit, speed and memory use) sets the ranking. Llama 4 Scout 17B-16EMeta · 109B (17.0B active) Q3_K_M 59.3 GB 11.1 tok/s Runs welloffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
60Good quality (B) · 60 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 55/100 (quality, fit, speed and memory use) sets the ranking. Devstral 2 (123B)Mistral AI · 125B Q2_K 57.4 GB 2.3 tok/s Very slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
57Good quality (B) · 57 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 55/100 (quality, fit, speed and memory use) sets the ranking. Command A Reasoning (111B)Cohere · 111B Q2_K 49.6 GB 3.2 tok/s Slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
56Good quality (B) · 56 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 54/100 (quality, fit, speed and memory use) sets the ranking. Command A (111B)Cohere · 111B Q2_K 49.6 GB 3.2 tok/s Slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
56Good quality (B) · 56 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 53/100 (quality, fit, speed and memory use) sets the ranking. Mistral Large 2.1 (2411)Mistral AI · 123B Q2_K 56.5 GB 2.4 tok/s Very slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
56Good quality (B) · 56 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 53/100 (quality, fit, speed and memory use) sets the ranking. Mistral Medium 3.5 (128B)Mistral AI · 128B Q2_K 58.5 GB 2.2 tok/s Very slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
54Fair quality (C) · 54 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 53/100 (quality, fit, speed and memory use) sets the ranking. Mistral Large 2 (123B)Mistral AI · 123B Q2_K 56.6 GB 2.4 tok/s Very slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
49Fair quality (C) · 49 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 51/100 (quality, fit, speed and memory use) sets the ranking. Command R+ (08-2024) 104BCohere · 104B Q2_K 48.2 GB 3.6 tok/s Slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
44Fair quality (C) · 44 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 49/100 (quality, fit, speed and memory use) sets the ranking. Command R+ 104B (04-2024)Cohere · 104B Q2_K 48.2 GB 3.6 tok/s Slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
53Fair quality (C) · 53 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 48/100 (quality, fit, speed and memory use) sets the ranking. dots.llm1 Instructrednote · 143B (14.0B active) Q2_K 71.0 GB 5.9 tok/s Slowheavy RAM offload NVIDIA PC (card + RAM) + Mac + Pi / ARM boardPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
36Basic quality (D) · 36 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 45/100 (quality, fit, speed and memory use) sets the ranking. Qwen1.5 110B ChatAlibaba · 111B Q2_K 51.6 GB 3.0 tok/s Very slowoffloads to RAM NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC Needs the clusterWon’t run on any one of them
38Basic quality (D) · 38 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 42/100 (quality, fit, speed and memory use) sets the ranking. Mixtral 8x22B InstructMistral AI · 141B (39.0B active) Q2_K 62.6 GB 6.2 tok/s Slowheavy RAM offload NVIDIA PC (card + RAM) + MacPipeline split llama.cpp RPC 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 88/100 (quality, fit, speed and memory use) sets the ranking. Kolibri 1 (78B-A3.5B)Aleph Alpha · 78B (3.5B active) Q3_K_M 42.9 GB 45.5 tok/s Runs welltight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Runs well on Mac, 95.9 tok/s
72Great quality (A) · 72 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. Qwen3 Coder Next (80B-A3B)Alibaba · 80B (3.0B active) Q3_K_M 43.7 GB 48.1 tok/s Runs welltight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Runs well on Mac, 106 tok/s
77Great quality (A) · 77 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. Nemotron 3.5 Bolt 74B-A9BNVIDIA · 75B (9.0B active) Q3_K_M 41.2 GB 27.3 tok/s Runs welltight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Runs well on Mac, 45.2 tok/s
77Great quality (A) · 77 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 79/100 (quality, fit, speed and memory use) sets the ranking. Llama 3.3 Nemotron Super 49B v1.5NVIDIA · 50B Q5_K_M 42.6 GB 4.7 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 7.5 tok/s
67Good quality (B) · 67 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking. Kimi Dev 72BMoonshot AI · 73B Q3_K_M 42.8 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
62Good quality (B) · 62 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 72/100 (quality, fit, speed and memory use) sets the ranking. Hermes 4 70BNous Research · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
62Good quality (B) · 62 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 72/100 (quality, fit, speed and memory use) sets the ranking. Llama 3.3 70B InstructMeta · 71B Q3_K_M 41.8 GB 5.1 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
61Good quality (B) · 61 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking. Qwen2.5 72B InstructAlibaba · 73B Q3_K_M 42.8 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
61Good quality (B) · 61 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking. Qwen2.5-VL 72B InstructAlibaba · 73B Q3_K_M 43.1 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
60Good quality (B) · 60 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek R1 Distill Llama 70BDeepSeek · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
60Good quality (B) · 60 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking. Tülu 3 70BAi2 · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
58Good quality (B) · 58 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 70/100 (quality, fit, speed and memory use) sets the ranking. Llama 3.1 70B InstructMeta · 71B Q3_K_M 41.8 GB 5.1 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
57Good quality (B) · 57 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 70/100 (quality, fit, speed and memory use) sets the ranking. Hermes 3 (Llama 3.1 70B)Nous Research · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
57Good quality (B) · 57 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 69/100 (quality, fit, speed and memory use) sets the ranking. QVQ 72B PreviewAlibaba · 73B Q3_K_M 43.1 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
55Good quality (B) · 55 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 69/100 (quality, fit, speed and memory use) sets the ranking. Llama 3.1 Nemotron 70B InstructNVIDIA · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
53Fair quality (C) · 53 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking. Qwen2 72B InstructAlibaba · 73B Q3_K_M 42.8 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
53Fair quality (C) · 53 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking. Qwen2-VL 72B InstructAlibaba · 73B Q3_K_M 43.1 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
50Fair quality (C) · 50 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking. Apertus 70B InstructSwiss AI · 71B Q3_K_M 41.8 GB 5.1 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
48Fair quality (C) · 48 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking. Llama 3 70B InstructMeta · 71B Q3_K_M 41.7 GB 5.2 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.7 tok/s
48Fair quality (C) · 48 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking. Qwen2.5-Math 72B InstructAlibaba · 73B Q3_K_M 41.5 GB 5.0 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.5 tok/s
36Basic quality (D) · 36 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 61/100 (quality, fit, speed and memory use) sets the ranking. Code Llama 70B InstructMeta · 69B Q3_K_M 39.6 GB 5.3 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 6.9 tok/s
33Basic quality (D) · 33 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 59/100 (quality, fit, speed and memory use) sets the ranking. DeepSeek LLM 67B ChatDeepSeek · 67B Q3_K_M 39.1 GB 5.4 tok/s Slowtight fit NVIDIA PC + MacPipeline split llama.cpp RPC Better on the clusterAlone: Slow on Mac, 7.0 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?