Which machines run Qwen3-Coder 30B-A3B?
Qwen/Qwen3-Coder-30B-A3B-Instruct · 30.5B parameters, 3B active (MoE) · license see model card · on Hugging Face since 2025-07-31
Memory for weights (estimate)
Q4_K_M: ≈18.5 GB
Q8_0: ≈32.4 GB
Computed from parameter count × bits per weight. Plus context.
Fits at 4-bit
22 of 24 tracked devices
Measured results
16 public tok/s results
Speeds come from different engines and quants.
| Device | Memory | Q4_K_M | Q8_0 | tok/s (measured) | Quant, engine | $ per tok/s | € per tok/s |
|---|---|---|---|---|---|---|---|
| NVIDIA RTX 5090 32GB (new) card, needs host PC | 32GB | fits | doesn't fit | 259.9 | as run, llm-speed.com, llama.cpp, median of 1 chat-short runs | – | – |
| GMKtec EVO-X2 128GB / 1TB | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $37 | €35 |
| GMKtec EVO-X2 128GB / 2TB | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $39 | €36 |
| GMKtec EVO-X3 128GB / 2TB | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $40 | €38 |
| Minisforum MS-S1 MAX 128GB / 2TB | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $41 | €36 |
| Beelink GTR9 Pro 128GB / 2TB | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $46 | €41 |
| Framework Desktop 128GB (DIY) | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $36 | €41 |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395) | 128GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | – | – |
| GMKtec EVO-X2 64GB / 1TB | 64GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $23 | €21 |
| Minisforum MS-S1 MAX 64GB / 2TB | 64GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $27 | €24 |
| Framework Desktop 64GB (DIY) | 64GB | fits | fits | 94.64 | UD-Q4_K_XL, llama.cpp Vulkan llama-bench tg128 | $26 | €23 |
| NVIDIA DGX Spark (Founders Edition) | 128GB | fits | fits | 61.06 | Q8_0, llama.cpp CUDA llama-bench tg32 | $77 | €68 |
| ASUS Ascent GX10 (GB10) | 128GB | fits | fits | 61.06 | Q8_0, llama.cpp CUDA llama-bench tg32 | $98 | €87 |
| Dell Pro Max / Pro Precision with GB10, 128GB / 2TB | 128GB | fits | fits | 61.06 | Q8_0, llama.cpp CUDA llama-bench tg32 | $97 | €86 |
| MSI EdgeXpert MS-C931 (GB10) | 128GB | fits | fits | 61.06 | Q8_0, llama.cpp CUDA llama-bench tg32 | – | – |
| Lenovo ThinkStation PGX (GB10) | 128GB | fits | fits | 61.06 | Q8_0, llama.cpp CUDA llama-bench tg32 | – | – |
| Mac Studio M5 Ultra 80-core GPU, 256GB / 1TB | 256GB | fits | fits | no public result | – | – | |
| Mac Studio M5 Max 40-core GPU, 128GB / 512GB | 128GB | fits | fits | no public result | – | – | |
| MacBook Pro 16" M5 Max 40-core GPU, 128GB / 2TB | 128GB | fits | fits | no public result | – | – | |
| Mac Studio M5 Ultra 64-core GPU, 96GB / 1TB | 96GB | fits | fits | no public result | – | – | |
| Mac mini M5 Pro 16-core GPU, 64GB / 512GB | 64GB | fits | fits | no public result | – | – | |
| Mac mini M5 Pro 20-core GPU, 64GB / 512GB | 64GB | fits | fits | no public result | – | – | |
| NVIDIA RTX 4090 24GB (used) card, needs host PC | 24GB | doesn't fit | doesn't fit | – | – | – | |
| NVIDIA RTX 3090 24GB (used) card, needs host PC | 24GB | doesn't fit | doesn't fit | – | – | – |
Fit is an estimate: weights × 1.15 + 2 GB must fit in usable memory (VRAM − 1 GB for cards, 85% of unified memory for machines). tok/s is text generation from the linked public source for that chip; "no public result" means we found none, never a guess. Why this model is listed: MoE coding model, 3B active.