← CanIRunLocal · All hardware · All models
Alibaba · 72.7B parameters · chat.
Ollama weight files: Q4_K_M 47 GB (qwen2.5:72b); Q5_K_M 54 GB (qwen2.5:72b-instruct-q5_K_M); Q8_0 77 GB (qwen2.5:72b-instruct-q8_0).
Yes on 25 machines in this list. The smallest is M1 Max 64GB: Q4_K_M, about 45 GB, roughly 8 tokens per second, via LM Studio. lms get mlx-community/Qwen2.5-72B-Instruct-4bit
ollama run qwen2.5:72b-instruct-q8_0ollama run qwen2.5:72b-instruct-q8_0ollama run qwen2.5:72b-instruct-q8_0ollama run qwen2.5:72b-instruct-q8_0lms get bartowski/Qwen2.5-72B-Instruct-GGUF@Q8_0llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 -c 8192vllm serve RedHatAI/Qwen2.5-72B-Instruct-FP8-dynamic --max-model-len 8192 --gpu-memory-utilization 0.9| Hardware | Runtime | Quant | Memory | Fit | Speed | Command |
|---|---|---|---|---|---|---|
| M1 Max 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~8 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| M2 Max 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~8 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| M3 Max 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~8 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| M4 Max 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~11 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| Mac Studio M4 Max (40c GPU) 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~11 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| Mac Studio M5 Max (40c GPU) 64GB | LM Studio | MLX 4bit | 45 GB | fits in fast memory | ~12 tok/s | lms get mlx-community/Qwen2.5-72B-Instruct-4bit |
| DGX Spark 64GB (ASUS, Dell, HP, MSI, Acer, Gigabyte) | Ollama | Q4_K_M | 51.1 GB | fits in fast memory | ~4 tok/s | ollama run qwen2.5:72b |
| Mac Studio M3 Ultra 96GB | Ollama | Q5_K_M | 58.1 GB | fits in fast memory | ~12 tok/s | ollama run qwen2.5:72b-instruct-q5_K_M |
| Mac Studio M5 Ultra 96GB | Ollama | Q5_K_M | 58.1 GB | fits in fast memory | ~18 tok/s | ollama run qwen2.5:72b-instruct-q5_K_M |
| M3 Max 128GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~4 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| M4 Max 128GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~6 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| Mac Studio M4 Max (40c GPU) 128GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~6 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| Mac Studio M5 Max (40c GPU) 128GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~6 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte) | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~3 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| M2 Ultra 192GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~8 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| 2x Mac Studio M5 Max (40c GPU) 128GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~5 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node |
| Mac Studio M3 Ultra 256GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~9 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| Mac Studio M5 Ultra (64c GPU) 256GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~12 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| Mac Studio M5 Ultra (80c GPU) 256GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~12 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| 2x DGX Spark 128GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~2 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node |
| 4x Mac Studio M5 Max (40c GPU) 128GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~5 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node |
| 2x Mac Studio M5 Ultra (80c GPU) 256GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~11 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052 -c 8192 # run rpc-server on each other node |
| Mac Studio M3 Ultra (80c GPU) 512GB | Ollama | Q8_0 | 81.1 GB | fits in fast memory | ~9 tok/s | ollama run qwen2.5:72b-instruct-q8_0 |
| 4x DGX Spark 128GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~2 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node |
| 4x Mac Studio M5 Ultra (80c GPU) 256GB | llama.cpp | Q8_0 GGUF | 81.4 GB | fits in fast memory | ~9 tok/s | llama-server -hf bartowski/Qwen2.5-72B-Instruct-GGUF:Q8_0 --rpc node2:50052,node3:50052,node4:50052 -c 8192 # run rpc-server on each other node |
Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-16.
A project by OnDevice Labs · llms.txt