Fastest rig for Qwen3-Coder-30B-A3B (local)
Qwen3-Coder-30B-A3B is a 30B mixture-of-experts with ~3B active — it decodes like a small model but codes like a big one. Here's the fastest decode tok/s submitted across every GPU and Apple Silicon tier.
The RTX 5090 is the fastest measured rig for Qwen3-Coder-30B-A3B at 259.9 tok/s (llama.cpp, r_c7qyvvmmsv1). A 24GB RTX 4090 measured 145.2 tok/s (Ollama Q4_K_M, r_wjq32z47vlp), followed by the M4 Max 128GB at 113.3 tok/s (Ollama Q4_K_M, r_roktphpc--8) and the M3 Ultra at 112.2 tok/s (MLX 4-bit, r_fpsca03u2o_). All four are the single-request chat-short workload; backends differ.
Fastest submitted run for Qwen3-Coder-30B-A3B: RTX 5090 (32GB) at 304.9tok/s.
Qwen3-Coder-30B-A3B keeps only ~3B of its 30B parameters active per token, so it fits at 4-bit on a single 24 GB GPU or any Apple Silicon Max/Ultra and decodes far faster than a dense 30B — memory bandwidth, not raw size, sets the speed. That's why a consumer GPU and an Apple Ultra can land surprisingly close. Below is every rig we have a real submitted run for, ranked by decode tok/s, each row linking the signed run. If your config isn't here yet, run the suite and submit — the row appears next refresh.
Submitted benchmarks
| Hardware | decode tok/s | Workload | Run |
|---|---|---|---|
| RTX 5090 (32GB) | 304.9tok/s | chat-short | r_vsype4pm1ko |
| RTX 4090 (24GB) | 232.5tok/s | chat-short | r_1k8942tw3zf |
| RTX 3090 (24GB) | 199.7tok/s | chat-short | r_0dh1kzbv4h0 |
| M3 Ultra (60-core GPU) | 112.2tok/s | chat-short | r_fpsca03u2o_ |
See also: All hardware · All models · Methodology