Qwen3-Coder 30B-A3B VRAM Requirements
Qwen3-Coder 30B-A3B needs about 18.8 GB of VRAM at Q4_K_M with an 8K context: 17.2 GB for weights, 0.8 GB for the KV cache and 0.8 GB of runtime overhead. At Q8_0 it needs 31.8 GB and at full FP16 precision 58.4 GB. The smallest GPU that runs it fully is the Radeon RX 7900 XT 20GB.
18.8 GB VRAM
Qwen3-Coder 30B-A3B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 17.2 GB
- KV cache 0.8 GB
- Runtime overhead 0.8 GB
Mixture of experts: all 30.5B parameters need memory, but only 3.3B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GB60% on GPU~30 tok/s
- RTX 4060 Ti 16GB80% on GPU~38 tok/s
- RTX 3090 24GBFits~114 tok/s
- Radeon RX 7900 XTX 24GBFits~107 tok/s
- RTX 4090 24GBFits~118 tok/s
- RTX 5090 32GBFits~146 tok/s
- RTX 6000 Ada 48GBFits~116 tok/s
- A100 80GBFits~151 tok/s
- H100 80GB SXMFits~169 tok/s
- Ryzen AI Max+ 395 128GBFits~49 tok/s
- Apple M4 Max 128GBFits~86 tok/s
- Apple M5 Max 128GBFits~92 tok/s
- RTX PRO 6000 Blackwell 96GBFits~146 tok/s
- NVIDIA DGX Spark 128GBFits~51 tok/s
- Apple M3 Ultra 512GBFits~107 tok/s
Qwen3-Coder 30B-A3B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 17.2 GB | 18.8 GB | 21.0 GB | 20 GB |
| Q5_K_M | 5.69 | 20.2 GB | 21.8 GB | 24.0 GB | 24 GB |
| Q6_K | 6.59 | 23.4 GB | 25.0 GB | 27.2 GB | 32 GB |
| Q8_0 | 8.50 | 30.2 GB | 31.8 GB | 34.0 GB | 48 GB |
| FP16 / BF16 | 16.00 | 56.8 GB | 58.4 GB | 60.6 GB | 64 GB |
| Q3_K_M | 3.91 | 13.9 GB | 15.5 GB | 17.7 GB | 20 GB |
| Q2_K | 3.00 | 10.7 GB | 12.2 GB | 14.5 GB | 16 GB |
VRAM by context length
At Q4_K_M. The KV cache grows linearly with context: 0.09 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.4 GB | 0.2 GB | 18.4 GB | 18.2 GB |
| 8K tokens | 0.8 GB | 0.4 GB | 18.8 GB | 18.4 GB |
| 32K tokens | 3.0 GB | 1.6 GB | 21.0 GB | 19.6 GB |
| 128K tokens | 12.0 GB | 6.4 GB | 30.0 GB | 24.4 GB |
| 256K tokens | 24.0 GB | 12.8 GB | 42.0 GB | 30.8 GB |
GPUs that can run Qwen3-Coder 30B-A3B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Radeon RX 7900 XT 20GB | 20.0 GB | Fits | Q4_K_M | 8K | ~97 tok/s |
| Apple M5 32GB | 21.3 GB | Fits | Q4_K_M | 21K | ~35 tok/s |
| NVIDIA L4 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~58 tok/s |
| Tesla P40 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~57 tok/s |
| Apple M4 Max 36GB | 24.0 GB | Fits | Q5_K_M | 48K | ~72 tok/s |
| Intel Arc Pro B60 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~65 tok/s |
| RTX 3090 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~114 tok/s |
| Radeon RX 7900 XTX 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~107 tok/s |
| RTX 3090 Ti 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~118 tok/s |
| RTX 4090 24GB | 24.0 GB | Fits | Q5_K_M | 48K | ~118 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | Fits | Q6_K | 128K | ~86 tok/s |
| RTX 5090 32GB | 32.0 GB | Fits | Q6_K | 128K | ~146 tok/s |
| Apple M4 Pro 48GB | 36.0 GB | Fits | Q8_0 | 168K | ~55 tok/s |
| A100 40GB | 40.0 GB | Fits | Q8_0 | 208K | ~139 tok/s |
| RTX 4060 Ti 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~38 tok/s |
| RTX 4070 Ti Super 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~48 tok/s |
| RTX 4080 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~49 tok/s |
| RTX 4080 Super 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~49 tok/s |
| RTX 5060 Ti 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~44 tok/s |
| RTX 5070 Ti 16GB | 16.0 GB | 80% on GPU | Q2_K | – | ~51 tok/s |
About Qwen3-Coder 30B-A3B
The local-friendly coding model in the Qwen3-Coder line, built on the Qwen 3 30B-A3B architecture with a 256K native context. It is popular with coding agents such as Cline, Roo Code and Continue because it is fast. Agents use long contexts, so the KV cache, not the weights, is usually what runs out first.
| Spec | Value |
|---|---|
| Parameters | 30.5B total, 3.3B active (MoE) |
| Layers | 48 |
| Hidden size | 2048 |
| Attention heads / KV heads | 32 / 4 |
| Head dimension | 128 |
| Vocabulary | 151,936 |
| Max context (used here) | 256K tokens |
| KV cache per token (FP16) | 96 KB |
| Released | 2025-07 |
| License | Apache 2.0 |
Qwen3-Coder 30B-A3B VRAM: common questions
How much VRAM does Qwen3-Coder 30B need for a 128K context?
About 12 GB of KV cache in FP16 on top of roughly 17 GB of Q4_K_M weights, so about 30 GB. A Q8_0 cache brings that to about 25 GB.
How much VRAM does Qwen3-Coder 30B-A3B need?
About 18.8 GB at Q4_K_M with an 8K context, 31.8 GB at Q8_0 and 58.4 GB at FP16. Longer contexts add 0.09 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Qwen3-Coder 30B-A3B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the Radeon RX 7900 XT 20GB (20.0 GB usable).
Does Qwen3-Coder 30B-A3B need VRAM for all 30.5B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 3.3B parameters are read per token, so it generates text about as fast as a 3B dense model.