Alibaba Qwen · 30.5B total / 3.3B active · MoE · fast chat & coding
What hardware do you need to run Qwen3 30B-A3B locally?
You need about 21 GB of GPU or unified memory to run Qwen3 30B-A3B at Q4_K_M (4-bit), or 36 GB at Q8_0 (8-bit). 17 of the 18 rigs we track can hold it. The cheapest is the GMKtec EVO-X2 (64 GB) at $2,200, at an estimated ~67 tok/s. The fastest is the NVIDIA GeForce RTX 5090 Founders Edition at an estimated ~260 tok/s. Prices & stock checked 2026-10-24. Prices change daily and some retailers block our checker, so always confirm on the store page.
Memory needed
| Q4_K_M (4-bit) | Q8_0 (8-bit) | |
|---|---|---|
| Memory needed (est.) | ~21 GB | ~36 GB |
| Weights only | ~18 GB | ~32 GB |
| Read per token | ~2.0 GB | ~3.5 GB |
| Rigs that fit | 17 of 18 | 14 of 18 |
It's a mixture-of-experts model: all 30.5B parameters have to fit in memory, but only 3.3B are read per token, so it runs much faster than a dense model of the same size.
A mixture-of-experts model with 30.5B total parameters and 3.3B active per token (128 experts, 8 active), from the April 2025 Qwen3 release. Model card
17 rigs that run Qwen3 30B-A3B
Sorted by today's lowest price. Quant is what we'd use on that machine: Q8_0 when it fits comfortably and still runs 10+ tok/s, otherwise Q4_K_M.
| Rig | Price today | Quant | Memory used | Est. tokens/sec | Buy |
|---|---|---|---|---|---|
| GMKtec EVO-X264 GB · 256 GB/s | $2,200GMKtec store · 2026-10-24 | Q8_0 | ~36 of 48 GB | ~45 est. | Check price |
| Mac Studio M5 Max36 GB · 460 GB/s | $2,499Apple Store · 2026-10-24 | Q4_K_M | ~21 of 27 GB | ~99 est. | Check price |
| Mac Studio M5 Max48 GB · 614 GB/s | $3,099Apple Store · 2026-10-24 | Q4_K_M | ~21 of 36 GB | ~112 est. | Check price |
| GMKtec EVO-X2128 GB · 256 GB/s | $3,500GMKtec store · 2026-10-24 | Q8_0 | ~36 of 96 GB | ~45 est. | Check price |
| Minisforum MS-S1 Max128 GB · 256 GB/s | $3,919Minisforum store · 2026-10-24 | Q8_0 | ~36 of 96 GB | ~45 est. | Check price |
| Beelink GTR9 Pro128 GB · 256 GB/s | $4,349Beelink store · 2026-10-24 | Q8_0 | ~36 of 96 GB | ~45 est. | Check price |
| Framework Desktop DIY128 GB · 256 GB/s | $4,839Framework store · 2026-10-24 | Q8_0 | ~36 of 96 GB | ~45 est. | Check price |
| Gigabyte GeForce RTX 4090 Windforce V224 GB · 1,008 GB/s (card only, needs a host PC) | $4,899Newegg · 2026-10-24 | Q4_K_M | ~21 of 22 GB | ~204 est. | Check price |
| Mac Studio M5 Ultra96 GB · 1,200 GB/s | $5,499Apple Store · 2026-10-24 | Q8_0 | ~36 of 72 GB | ~100 est. | Check price |
| Lenovo ThinkStation PGX128 GB · 273 GB/s | $5,859B&H Photo · 2026-10-24 | Q8_0 | ~36 of 109 GB | ~50 est. | Check price |
| ASUS Ascent GX10128 GB · 273 GB/s | $5,999B&H Photo · 2026-10-24 | Q8_0 | ~36 of 109 GB | ~50 est. | Check price |
| NVIDIA DGX Spark128 GB · 273 GB/s | $6,950NVIDIA Marketplace · 2026-10-24 | Q8_0 | ~36 of 109 GB | ~50 est. | Check price |
| HP ZGX Nano G1n128 GB · 273 GB/s | $7,699B&H Photo · 2026-10-24 | Q8_0 | ~36 of 109 GB | ~50 est. | Check price |
| Gigabyte AI TOP ATOM128 GB · 273 GB/s | $8,499B&H Photo · 2026-10-24 | Q8_0 | ~36 of 109 GB | ~50 est. | Check price |
| NVIDIA GeForce RTX 5090 Founders Edition32 GB · 1,792 GB/s (card only, needs a host PC) | $8,500Newegg · 2026-10-24 | Q4_K_M | ~21 of 30 GB | ~260 est. | Check price |
| PNY NVIDIA RTX PRO 6000 Blackwell96 GB · 1,792 GB/s (card only, needs a host PC) | $17,899B&H Photo · 2026-10-24 | Q8_0 | ~36 of 89 GB | ~206 est. | Check price |
| Mac Studio M5 Ultra256 GB · 1,200 GB/s | No price seen | Q8_0 | ~36 of 192 GB | ~100 est. | Check price |
Tokens/sec are estimates from our bandwidth formula (one user, ~8K context, llama.cpp-class runtime), not benchmarks. GPU prices are for the card only. Prices change daily, so confirm on the store page. How we estimate
Measured speeds from public benchmarks
Single-user decode speeds people have published for this model. These are measured, not our estimates.
| Hardware | Model / run | Measured tok/s | Source |
|---|---|---|---|
| Strix Halo256 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp Vulkan tg128 | 85.5 | source |
| Strix Halo256 GB/s | Qwen3 30B-A3B UD-Q4_K_XLllama.cpp Vulkan tg128 | 75.3 | source |
| RTX 40901,008 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp tg128 (fusion off) | 198.4 | source |
| RTX 50901,792 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp + GGML_CUDA_GRAPH_OPT=1 | 352.1 | source |
From the benchmark set we use to calibrate our speed formula. All 52 runs
Every rig we track, 4-bit and 8-bit
| Machine | Price | Q4_K_M | Q8_0 |
|---|---|---|---|
| Mac mini M5 Pro15-core CPU / 16-core GPU, 24GB | $1,699 | Won't fitneeds ~21 GB | Won't fitneeds ~36 GB |
| GMKtec EVO-X2Ryzen AI Max+ 395, 64GB / 1TB | $2,200 | Fits~67 tok/s21 of 48 GB | Fits~45 tok/s36 of 48 GB |
| Mac Studio M5 Max18-core CPU / 32-core GPU, 36GB | $2,499 | Fits~99 tok/s21 of 27 GB | Won't fitneeds ~36 GB |
| Mac Studio M5 Max18-core CPU / 40-core GPU, 48GB | $3,099 | Fits~112 tok/s21 of 36 GB | Tight fit~86 tok/s36 of 36 GB |
| GMKtec EVO-X2Ryzen AI Max+ 395, 128GB / 1TB | $3,500 | Fits~67 tok/s21 of 96 GB | Fits~45 tok/s36 of 96 GB |
| Minisforum MS-S1 MaxRyzen AI Max+ 395, 128GB / 2TB | $3,919 | Fits~67 tok/s21 of 96 GB | Fits~45 tok/s36 of 96 GB |
| Beelink GTR9 ProRyzen AI Max+ 395, 128GB / 2TB | $4,349 | Fits~67 tok/s21 of 96 GB | Fits~45 tok/s36 of 96 GB |
| Framework Desktop DIYRyzen AI Max+ 395, 128GB | $4,839 | Fits~67 tok/s21 of 96 GB | Fits~45 tok/s36 of 96 GB |
| Gigabyte GeForce RTX 4090 Windforce V224GB card | $4,899 | Tight fit~204 tok/s21 of 22 GB | Won't fitneeds ~36 GB |
| Mac Studio M5 Ultra30-core CPU / 64-core GPU, 96GB | $5,499 | Fits~118 tok/s21 of 72 GB | Fits~100 tok/s36 of 72 GB |
| Lenovo ThinkStation PGX1TB | $5,859 | Fits~77 tok/s21 of 109 GB | Fits~50 tok/s36 of 109 GB |
| ASUS Ascent GX101TB | $5,999 | Fits~77 tok/s21 of 109 GB | Fits~50 tok/s36 of 109 GB |
| NVIDIA DGX SparkFounders Edition, 4TB | $6,950 | Fits~77 tok/s21 of 109 GB | Fits~50 tok/s36 of 109 GB |
| HP ZGX Nano G1n4TB | $7,699 | Fits~77 tok/s21 of 109 GB | Fits~50 tok/s36 of 109 GB |
| Gigabyte AI TOP ATOM4TB | $8,499 | Fits~77 tok/s21 of 109 GB | Fits~50 tok/s36 of 109 GB |
| NVIDIA GeForce RTX 5090 Founders Edition32GB card | $8,500 | Fits~260 tok/s21 of 30 GB | Won't fitneeds ~36 GB |
| PNY NVIDIA RTX PRO 6000 Blackwell96GB card | $17,899 | Fits~260 tok/s21 of 89 GB | Fits~206 tok/s36 of 89 GB |
| Mac Studio M5 Ultra256GB | — | Fits~118 tok/s21 of 192 GB | Fits~100 tok/s36 of 192 GB |
Fit and tokens/sec are estimates. "Tight fit" means over 90% of usable memory: it loads, but leaves little room for long context.
FAQ
How much memory do I need to run Qwen3 30B-A3B locally?
About 21 GB at Q4_K_M (4-bit) and 36 GB at Q8_0 (8-bit). That's the weights plus runtime overhead and roughly 8K tokens of context for one user. Longer context needs more. These are estimates.
What's the cheapest hardware that runs Qwen3 30B-A3B?
Of the rigs we track, the GMKtec EVO-X2 64 GB at $2,200 (GMKtec store, checked 2026-10-24). We estimate about 67 tokens/sec at 4-bit.
How fast does Qwen3 30B-A3B run locally?
We estimate about 67 to 260 tokens/sec at 4-bit across the rigs that fit, fastest on the NVIDIA GeForce RTX 5090 Founders Edition. Published runs we've collected range from 75.3 to 352.1 tok/s.
Can I run Qwen3 30B-A3B on a single gaming GPU like the RTX 4090 or 5090?
Yes. NVIDIA GeForce RTX 5090 Founders Edition (32 GB): ~260 tok/s estimated at 4-bit; Gigabyte GeForce RTX 4090 Windforce V2 (24 GB): ~204 tok/s estimated at 4-bit.
Can a Mac run Qwen3 30B-A3B?
Yes. The cheapest Mac we track that fits it is the Mac Studio M5 Max 36 GB at $2,499, at an estimated ~99 tok/s. macOS lets the GPU use about 75% of memory by default.
Are these speeds measured or estimated?
The table shows estimates from memory bandwidth, calibrated against 43 public llama.cpp/MLX benchmarks (median error about 5%). The measured-speeds table shows published runs with links.