can it run
Yes, but close to the edge. 24GB against a 24GB working set: it runs, with little room for long context windows or other processes.
weights 18.6GB + KV cache 0.38GB at 4k context vs ~22GB usable → tight. 19.0 of 22 GB usable — keep context modest
computed roofline: 273 GB/s × 0.25–0.65 efficiency window / 1.9GB (Q4_K_M) — bands, never points.
MoE: file is 18.6GB (that decides fit) but only ~3B of experts activate per token (~1.9GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math
| model file | 18.6GB (GGUF Q4_K_M) |
| minimum memory | 24GB working set |
| this machine | 24GB unified LPDDR5X (18-core CPU / 5,120-core GPU base config) |
| params | 30B-A3B MoE |
| license | apache-2.0 |
| price | $2599.99 |
easy run: ollama pull qwen3:30b-a3b
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
→ rtx-5090 — the next step up from this machine's GPU