Qwen3-Coder 30B-A3B: VRAM requirements & speed
Qwen3-Coder 30B-A3B (30.53B MoE, 3.3B active, apache-2.0). Native context 256k, 1M with YaRN. Smallest GGUF: 18.56 GB.
Memory needed by quantization and context
GGUF file size plus f16 KV cache (and fixed overhead of ~1 GB per GPU not shown). Source: unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF.
| Quant | File | +8k ctx | +32k ctx | +128k ctx | +256k ctx | +1M ctx |
|---|---|---|---|---|---|---|
| Q4_K_M | 18.56 GB | 19.4 GB | 21.8 GB | 31.4 GB | 44.3 GB | 116.9 GB |
| Q5_K_M | 21.73 GB | 22.5 GB | 25.0 GB | 34.6 GB | 47.5 GB | 120.0 GB |
| Q6_K | 25.09 GB | 25.9 GB | 28.3 GB | 38.0 GB | 50.9 GB | 123.4 GB |
| Q8_0 | 32.48 GB | 33.3 GB | 35.7 GB | 45.4 GB | 58.2 GB | 130.8 GB |
KV cache size by context length
Computed from the model's attention layout.
| Context | f16 | q8_0 | q4_0 |
|---|---|---|---|
| 8k | 0.81 GB | 0.43 GB | 0.23 GB |
| 32k | 3.22 GB | 1.71 GB | 0.91 GB |
| 128k | 12.88 GB | 6.85 GB | 3.62 GB |
| 256k | 25.77 GB | 13.69 GB | 7.25 GB |
| 1M YaRN | 98.30 GB | 52.22 GB | 27.65 GB |
Which hardware runs Qwen3-Coder 30B-A3B? · 16k context
| Hardware | Quant | Fit | Memory | Tok/s | Speed source |
|---|---|---|---|---|---|
| RTX 5090 32GB | Q6_K | fits fully | 26.7 GB | 316 | scaled from measured |
| RTX 4090 24GB | Q4_K_M | fits fully | 20.2 GB | 266 | measured |
| RTX PRO 6000 96GB (bonus page) | Q8_0 | fits fully | 34.1 GB | 246 | estimated |
| AMD RX 7900 XTX 24GB | Q4_K_M | fits fully | 20.2 GB | 228 | estimated |
| RTX 3090 24GB | Q4_K_M | fits fully | 20.2 GB | 206 | measured |
| Mac Studio M5 Ultra 256GB | Q8_0 | fits fully | 34.1 GB | 169 | estimated |
| Mac Studio M5 Ultra 512GB (ships late Oct 2026) | Q8_0 | fits fully | 34.1 GB | 169 | estimated |
| 2x RTX 3090 (48GB total) | Q8_0 | fits fully | 34.1 GB | 143 | estimated |
| RTX 5080 16GB | Q4_K_M | MoE offload to CPU | 20.2 GB | 81 | estimated |
| RTX 5070 Ti 16GB | Q4_K_M | MoE offload to CPU | 20.2 GB | 79 | estimated |
| Mac Studio (2025) M3 Ultra 96GB | Q8_0 | fits fully | 34.1 GB | 73 | estimated |
| Mac Studio M5 Max 128GB | Q8_0 | fits fully | 34.1 GB | 65 | scaled from measured |
| MacBook Pro 14/16 M5 Max 128GB | Q8_0 | fits fully | 34.1 GB | 65 | scaled from measured |
| 16GB VRAM class (RTX 4060 Ti 16GB, 32GB RAM) | Q4_K_M | MoE offload to CPU | 20.2 GB | 63 | estimated |
| MacBook Pro 14/16 M5 Pro 48GB | Q6_K | fits fully | 26.7 GB | 62 | estimated |
| NVIDIA DGX Spark 128GB | Q8_0 | fits fully | 34.1 GB | 61 | measured |
| MacBook Pro 14/16 (2024) M4 Max 128GB | Q8_0 | fits fully | 34.1 GB | 58 | estimated |
| AMD Ryzen AI Max+ 395 (Strix Halo) 128GB | Q8_0 | fits fully | 34.1 GB | 55 | scaled from measured |
| RTX 5070 12GB | Q4_K_M | MoE offload to CPU | 20.2 GB | 48 | estimated |
| RTX 4070 12GB | Q4_K_M | MoE offload to CPU | 20.2 GB | 47 | estimated |
| Mac mini M6 32GB | Q4_K_M | fits fully | 20.2 GB | 42 | estimated |
| MacBook Air M5 32GB | Q4_K_M | fits fully | 20.2 GB | 39 | estimated |
| Mac mini (2024) M4 Pro 64GB | Q8_0 | fits fully | 34.1 GB | 35 | estimated |
| 8GB VRAM class (RTX 4060 8GB, 32GB RAM) | Q4_K_M | MoE offload to CPU | 20.2 GB | 34 | estimated |
| RTX 3060 12GB (32GB RAM) | Q4_K_M | MoE offload to CPU | 20.2 GB | 28 | estimated |
| Mac mini / iMac M4 16GB | - | does not fit | - | - |
Discrete GPUs are modelled with 32 GB of system RAM available for offload. Does not fit even with offload: Mac mini / iMac M4 16GB.
Links
Model card · GGUF files · Open in the advisor · last verified 2026-10-14
FAQ
How much VRAM does Qwen3-Coder 30B-A3B need?
The smallest quant (Q4_K_M) is 18.56 GB, so you need roughly 20 GB for an 8k context including KV cache and overhead. At 1M context the f16 KV cache alone is 98.3 GB.
Can I run Qwen3-Coder 30B-A3B on a RTX 5090 32GB?
Yes: on RTX 5090 32GB it fits fully as Q6_K at about 316 tok/s (scaled from measured). 18 of 26 hardware configurations fit it fully at 16k context.
What is the maximum context for Qwen3-Coder 30B-A3B?
256k tokens natively, up to 1M with YaRN as documented on the model card. Memory, not the model, is usually the limit: see the KV cache table above.
All models: model index · by memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB. Numbers use the same fit math as the advisor; sources and methods.