can it run
Only with compromises. 32GB is below the 33GB working set — expect a lower quant, offloading, or heavy swap. Possible, not pleasant.
weights 26GB + KV cache 0.5GB at 4k context vs ~32GB usable → tight. 26.5 of 32 GB usable — keep context modest
computed roofline: 897 GB/s × 0.46–0.76 efficiency window / 3.9GB (Q4_K_M) — bands, never points.
MoE: file is 26GB (that decides fit) but only ~1.05B of experts activate per token (~3.9GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math
| model file | 26GB (GGUF Q4_K_M) |
| minimum memory | 33GB working set |
| this machine | 32GB HBM2, 897 GB/s |
| params | 8x7B MoE |
| license | apache-2.0 |
| price | $645 |
easy run: ollama pull mixtral:8x7b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ Mac mini M5 Pro (64GB) — cheapest machine that runs it ($1669)
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine
→ every model the Tesla V100 32GB (used) can run — full list
→ RTX 3090 (24GB, used) — upgrade path machine