dexiio

What hardware runs Qwen3 Coder 30B A3B?

Qwen3 Coder 30B A3B is a 30.53B-parameter mixture-of-experts model with 3.35B active parameters (apache-2.0 license). Purpose-built agentic coder. Best local fill-in-the-middle and tool-calling under 70B; useless at small talk.

All figures assume an f16 KV cache, a 0.6 GB display reserve on the GPU, and 64 GB of DDR5 system RAM for the offload tiers. Tune these in the calculator.

Minimum VRAM by quant and context

QuantFile size8K16K32K64K128K
Q2_K *10.5 GB13.1 GB13.8 GB15.4 GB18.5 GB24.8 GB
Q3_K_M *13.7 GB16.3 GB17.1 GB18.6 GB21.8 GB28 GB
Q4_K_M17.3 GB19.9 GB20.6 GB22.2 GB25.3 GB31.6 GB
Q5_K_M20.2 GB22.8 GB23.6 GB25.2 GB28.3 GB34.5 GB
Q6_K23.4 GB26 GB26.7 GB28.3 GB31.4 GB37.7 GB
Q8_030.3 GB32.8 GB33.6 GB35.2 GB38.3 GB44.6 GB
IQ4_XS15.3 GB17.8 GB18.6 GB20.2 GB23.3 GB29.6 GB

Full-GPU figures: weights + f16 KV cache + overhead. * below our recommended floor of Q4_K_M.

GPU compatibility

GPU8K16K32K64K128K
NVIDIA GeForce RTX 5090Q6_KQ6_KQ6_KQ6_KQ4_K_M
NVIDIA GeForce RTX 5080Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5070 TiQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5070Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5060Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4090Q5_K_MQ5_K_MQ4_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 4080 SUPERQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4070 Ti SUPERQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4070Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 4060Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 3090Q5_K_MQ5_K_MQ4_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 3080 10GBQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 3070Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 3060 TiQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 3060 12GBQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 2080 TiQ8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce GTX 1080 TiQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 9070 XTQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 7900 XTXQ5_K_MQ5_K_MQ4_K_MIQ4_XSQ8_0
AMD Radeon RX 7900 XTQ4_K_MIQ4_XSQ8_0Q8_0Q8_0
AMD Radeon RX 7800 XTQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 7600 XTQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 6800 XTQ8_0Q8_0Q8_0Q8_0Q8_0
AMD Radeon RX 6700 XTQ8_0Q8_0Q8_0Q8_0Q8_0
Intel Arc B580Q8_0Q8_0Q8_0Q8_0Q8_0
Intel Arc A770 16GBQ8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 (16GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 Pro (32GB)Q5_K_MQ5_K_MQ4_K_MIQ4_XSQ8_0
Apple M2 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 Ultra (128GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 (16GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Pro (36GB)Q6_KQ6_KQ5_K_MQ4_K_MQ8_0
Apple M3 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Ultra (96GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 (16GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 Pro (48GB)Q8_0Q8_0Q8_0Q6_KQ5_K_M
Apple M4 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
CPU only / integrated graphicsQ8_0Q8_0Q8_0Q8_0Q8_0

Fits on GPUExpert offloadPartial offloadCPU only

Quant guidance

Our floor for Qwen3 Coder 30B A3B is Q4_K_M — below that, quality degrades faster than the VRAM savings are worth. Prefer the highest quant that still lands "Fits on GPU" or “Expert offload” at your context length in the table above.

Recommended run command

Q4_K_M at 32K on a NVIDIA GeForce RTX 5060-class GPU (Expert offload):

llama-server -m Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf -c 32768 --flash-attn -ngl 99 --n-cpu-moe 48

Frequently asked questions

How much VRAM does Qwen3 Coder 30B A3B need?

At Q4_K_M and 32K context, Qwen3 Coder 30B A3B needs about 22.2 GB of VRAM to run fully on GPU (weights + KV cache + overhead). As a mixture-of-experts model it can also run with far less VRAM via expert offload, keeping experts in system RAM.

What is the smallest GPU that can run Qwen3 Coder 30B A3B?

The NVIDIA GeForce RTX 5060 (8 GB) is the smallest GPU in our set that runs Qwen3 Coder 30B A3B well at 32K context, using expert offload with 64 GB of system RAM.

What quantization should I use for Qwen3 Coder 30B A3B?

We recommend Q4_K_M or higher. Q4_K_M weighs 17.3 GB (4.86 bits/weight); going below Q4_K_M costs noticeable quality on a model this size.

How long a context can Qwen3 Coder 30B A3B handle?

Qwen3 Coder 30B A3B supports up to 256K tokens. KV cache grows linearly with context.

Can I run Qwen3 Coder 30B A3B without a GPU?

Yes, at reduced speed: on CPU with 64 GB of DDR5 it manages roughly 14 tokens/sec at 8K context (Q8_0).

Related guides