RunLLMLocal
Open the checker

Compare · Models

Qwen2.5-Coder 32B vs Qwen3-Coder 30B-A3B: which runs better locally?

Qwen2.5-Coder 32B needs about 23.3 GB at Q4 with 8K context and fits fully on 1 of 40 GPUs. Qwen3-Coder 30B-A3B needs 20.6 GB and fits on 4. On cards that run both, Qwen3-Coder 30B-A3B is about 3.4x faster because it is a mixture-of-experts model.

VerdictQwen3-Coder 30B-A3B runs faster at a similar size

Qwen2.5-Coder 32B needs about 23.3 GB at Q4 with 8K context and fits fully on 1 of 40 GPUs. Qwen3-Coder 30B-A3B needs 20.6 GB and fits on 4. On cards that run both, Qwen3-Coder 30B-A3B is about 3.4x faster because it is a mixture-of-experts model.

Qwen2.5-Coder 32BCoding
Parameters32.8B
Memory at Q4, 8K23.3 GB
Memory at Q8, 8K40.3 GB
GPUs that run it fully1 / 40
Cheapest desktop cardRTX 5090 (32 GB)
Qwen3-Coder 30B-A3BCoding
Parameters30.5B (3.3B active)
Memory at Q4, 8K20.6 GB
Memory at Q8, 8K36.3 GB
GPUs that run it fully4 / 40
Cheapest desktop cardRTX 3090 / 3090 Ti (24 GB)

Each GPU at 8K context with the best quantization that fits. Offload assumes 32 GB of system RAM. Green marks the better result. Rows where the two differ come first.

GPUQwen2.5-Coder 32BQwen3-Coder 30B-A3B
RTX 4090 24 GBOffload~4.8 tok/sQ4_K_MRuns~130 tok/sQ4_K_M
RX 7900 XTX 24 GBOffload~4.6 tok/sQ4_K_MRuns~130 tok/sQ4_K_M
RTX 3090 / 3090 Ti 24 GBOffload~4.5 tok/sQ4_K_MRuns~120 tok/sQ4_K_M
RTX 5090 32 GBRuns~50 tok/sQ5_K_MRuns~170 tok/sQ6_K
RX 7900 XT 20 GBOffload~3.5 tok/sQ4_K_MOffload~15 tok/sQ4_K_M
RTX 5080 16 GBOffload~3.7 tok/sQ4_K_MOffload~17 tok/sQ4_K_M
RTX 5070 Ti 16 GBOffload~3.5 tok/sQ4_K_MOffload~16 tok/sQ4_K_M
RTX 4080 / 4080 Super 16 GBOffload~3.0 tok/sQ4_K_MOffload~14 tok/sQ4_K_M
RTX 4070 Ti Super 16 GBOffload~2.8 tok/sQ4_K_MOffload~13 tok/sQ4_K_M
RX 9070 XT 16 GBOffload~2.7 tok/sQ4_K_MOffload~13 tok/sQ4_K_M
RX 7800 XT 16 GBOffload~2.6 tok/sQ4_K_MOffload~13 tok/sQ4_K_M
Arc A770 16 GBOffload~2.4 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RX 6800 XT 16 GBOffload~2.3 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RTX 5060 Ti 16 GBOffload~2.1 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
RTX 4060 Ti 16 GBOffload~1.6 tok/sQ4_K_MOffload~7.9 tok/sQ4_K_M
RTX 5070 12 GBOffload~2.5 tok/sQ4_K_MOffload~13 tok/sQ4_K_M
RTX 4070 / 4070 Super 12 GBOffload~2.1 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
Arc B580 12 GBOffload~2.0 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RX 7700 XT 12 GBOffload~2.0 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RX 6700 XT 12 GBOffload~1.9 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
RTX 3060 12 GBOffload~1.8 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
GTX 1080 Ti 11 GBOffload~2.1 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RTX 3080 10 GBOffload~2.6 tok/sQ4_K_MOffload~14 tok/sQ4_K_M
Arc A750 8 GBOffload~2.0 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 5060 Ti 8 GBOffload~1.9 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 5060 8 GBOffload~1.9 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 3070 / 3070 Ti 8 GBOffload~1.9 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 3060 Ti 8 GBOffload~1.9 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 4060 Ti 8 GBOffload~1.7 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RX 7600 8 GBOffload~1.7 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RTX 4060 8 GBOffload~1.7 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
RTX 3050 8 GBOffload~1.6 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
GTX 1660 / 1660 Super 6 GBOffload~1.8 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
GTX 1650 4 GBOffload~2.0 tok/sQ3_K_MOffload~11 tok/sQ4_K_M
RTX 4090 Laptop 16 GBOffload~2.5 tok/sQ4_K_MOffload~12 tok/sQ4_K_M
RTX 4080 Laptop 12 GBOffload~2.0 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RTX 4070 Laptop 8 GBOffload~1.6 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
RTX 4060 Laptop 8 GBOffload~1.6 tok/sQ4_K_MOffload~10 tok/sQ4_K_M
RTX 3060 Laptop 6 GBOffload~1.8 tok/sQ4_K_MOffload~11 tok/sQ4_K_M
RTX 4050 Laptop 6 GBOffload~1.6 tok/sQ4_K_MOffload~11 tok/sQ4_K_M

Which should you run?

Qwen3-Coder 30B-A3B runs faster at a similar size. Qwen2.5-Coder 32B needs about 23.3 GB at Q4 with 8K context and fits fully on 1 of 40 GPUs. Qwen3-Coder 30B-A3B needs 20.6 GB and fits on 4. On cards that run both, Qwen3-Coder 30B-A3B is about 3.4x faster because it is a mixture-of-experts model.

Only Qwen3-Coder 30B-A3B fits on RTX 4090, RTX 3090 / 3090 Ti and RX 7900 XTX.

See every result on its own page: Qwen2.5-Coder 32B and Qwen3-Coder 30B-A3B.

Questions

Is Qwen2.5-Coder 32B or Qwen3-Coder 30B-A3B better?

Qwen3-Coder 30B-A3B runs faster at a similar size. Qwen2.5-Coder 32B needs about 23.3 GB at Q4 with 8K context and fits fully on 1 of 40 GPUs. Qwen3-Coder 30B-A3B needs 20.6 GB and fits on 4. On cards that run both, Qwen3-Coder 30B-A3B is about 3.4x faster because it is a mixture-of-experts model. The right pick depends on what you use it for: compare the strengths listed on each model's page.

Which needs less VRAM, Qwen2.5-Coder 32B or Qwen3-Coder 30B-A3B?

Qwen2.5-Coder 32B needs about 23 GB at Q4 with 8K context, Qwen3-Coder 30B-A3B about 21 GB. At Q8 it is 40 GB versus 36 GB.

Can I run Qwen2.5-Coder 32B and Qwen3-Coder 30B-A3B on a 12 GB GPU?

On an RTX 3060 12 GB, Qwen2.5-Coder 32B only runs with part of the model in system RAM, at a few tokens per second and Qwen3-Coder 30B-A3B only runs with part of the model in system RAM, at a few tokens per second.

More comparisons

Qwen3 14B vs Gemma 4 12BModel comparisonQwen3 8B vs Llama 3.1 8BModel comparisongpt-oss 20B vs Qwen3 30B-A3B (MoE)Model comparisonQwen3 30B-A3B (MoE) vs Qwen3 32BModel comparisonQwen3.6 35B-A3B (MoE) vs Qwen3 30B-A3B (MoE)Model comparisonGemma 4 31B vs Qwen3.6 27BModel comparison

Estimates from memory size and bandwidth at 8K context. Real speed varies with software, drivers and settings.

Copied