Models / Qwen3 30B-A3B
Best home rig for Qwen3 30B-A3B
Quick answer
How much VRAM does Qwen3 30B-A3B need?
The weights are about 18.7 GB at Q4_K_M (30.5B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Apple M5 Ultra (256 GB) at 162.2 tok/s.
What is the fastest measured hardware for Qwen3 30B-A3B?
Apple M5 Ultra: 162.2 tok/s generation (mlx-lm 0.32.2 (MLX 0.32.4), 4-bit MLX (mlx-community)), from https://bohemiansilicon.com/blog/two-mac-studio-m5-ultra-same-local-ai-performance/.
Last updated:
Legacy model30.5B total · 3.3B active3 measured resultsfastest plain decode: 162.2 tok/s model info ↗
Older model. For current picks see current models.
Real, measured generation speeds for Qwen3 30B-A3B on home hardware, fastest first. Different quants/engines aren't directly comparable.
| Build | Quant | Engine | tok/s | Prompt processing (tok/s) | Cost | Source |
|---|---|---|---|---|---|---|
| Mac Studio M5 Ultra 256 GB: Qwen3 30B-A3B 162.2 tok/s, gpt-oss 120B 121.4, Llama 3.3 70B 25.2 on MLX, two units within 0.35% (Bohemian Silicon) | 4-bit MLX (mlx-community) | mlx-lm 0.32.2 (MLX 0.32.4) | 162.2 | 7,646 @8K · est. TTFT @8k ≈1.1 s* | bohemiansilicon.com ↗ | |
| NVIDIA T1000 4 GB + i7-12700K: Qwen3 30B-A3B IQ3_XXS at 18.7 tok/s with experts in RAM, 27.1 with a 0.6B draft model | Unsloth UD-IQ3_XXS (12.0 GB) | llama.cpp llama-server b12121+ (CUDA) | 27.11 ⚡ | – | github.com ↗ | |
| NVIDIA T1000 4 GB + i7-12700K: Qwen3 30B-A3B IQ3_XXS at 18.7 tok/s with experts in RAM, 27.1 with a 0.6B draft model | Unsloth UD-IQ3_XXS (12.0 GB) | llama.cpp llama-server b12121+ (CUDA) | 18.68 | – | github.com ↗ |
Prompt processing = prefill speed reported by the source, @ prompt length. * est. TTFT @8k is calculated from measured prefill, not measured. Why it matters.
Memory math: how much VRAM / RAM does Qwen3 30B-A3B need?
Qwen3 30B-A3B has 30.5B parameters, about 3.3B active per token (source: Qwen model card: 30.5B total / 3.3B activated MoE (2025)). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.
| Format | Bits/weight | Weights | Smallest common memory size that holds the weights alone |
|---|---|---|---|
| Q8_0 | 8.50 | 32.4 GB | 48 GB |
| Q6_K | 6.56 | 25.0 GB | 32 GB |
| Q5_K_M | 5.70 | 21.7 GB | 24 GB |
| Q4_K_M | 4.89 | 18.7 GB | 24 GB |
| Q3_K_M | 4.00 | 15.2 GB | 16 GB |
| Q2_K | 3.16 | 12.0 GB | 16 GB |
Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.
Why it can still run on smaller GPUs: it is a mixture-of-experts model, so only about 3.3B of its 30.5B parameters are used for each generated token. Runtimes such as llama.cpp (--n-cpu-moe), ExLlamaV3 or Project Maya can keep most expert weights in system RAM and only the shared layers and hot experts in VRAM. Then you need VRAM + system RAM ≥ weights, and speed is limited mainly by RAM bandwidth.
Measured on setups with less GPU memory than the Q4_K_M weights (so part of the model lived in system RAM or was streamed from disk):
- NVIDIA T1000 4GB (4 GB) + 32 GB: 18.68 tok/s (Unsloth UD-IQ3_XXS (12.0 GB))
- NVIDIA T1000 4GB (4 GB) + 32 GB: 27.11 tok/s (Unsloth UD-IQ3_XXS (12.0 GB))
FAQ
How much VRAM does Qwen3 30B-A3B need?
The weights are about 18.7 GB at Q4_K_M (30.5B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Apple M5 Ultra (256 GB) at 162.2 tok/s.
What is the fastest measured hardware for Qwen3 30B-A3B?
Apple M5 Ultra: 162.2 tok/s generation (mlx-lm 0.32.2 (MLX 0.32.4), 4-bit MLX (mlx-community)), from https://bohemiansilicon.com/blog/two-mac-studio-m5-ultra-same-local-ai-performance/.
Related: what GPU do you need for 2026 models · all builds
Speed with and without MTP / DFlash on the same machines: measured speed-ups.