Compare · Models
gpt-oss 20B vs Qwen3 30B-A3B (MoE): which runs better locally?
gpt-oss 20B needs about 14.0 GB at Q4 with 8K context and fits fully on 16 of 40 GPUs. Qwen3 30B-A3B (MoE) needs 20.6 GB and fits on 4. On cards that run both, Qwen3 30B-A3B (MoE) is about 34% faster. Qwen3 30B-A3B (MoE) has more parameters (30.5B vs 21B), which usually means better answers if you have the memory.
gpt-oss 20B needs about 14.0 GB at Q4 with 8K context and fits fully on 16 of 40 GPUs. Qwen3 30B-A3B (MoE) needs 20.6 GB and fits on 4. On cards that run both, Qwen3 30B-A3B (MoE) is about 34% faster. Qwen3 30B-A3B (MoE) has more parameters (30.5B vs 21B), which usually means better answers if you have the memory.
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.
| GPU | gpt-oss 20B | Qwen3 30B-A3B (MoE) |
|---|---|---|
| RX 7900 XT 20 GB | Runs~89 tok/sQ5_K_M | Offload~15 tok/sQ4_K_M |
| RTX 5080 16 GB | Runs~130 tok/sQ4_K_M | Offload~17 tok/sQ4_K_M |
| RTX 5070 Ti 16 GB | Runs~120 tok/sQ4_K_M | Offload~16 tok/sQ4_K_M |
| RTX 4080 / 4080 Super 16 GB | Runs~98 tok/sQ4_K_M | Offload~14 tok/sQ4_K_M |
| RTX 4070 Ti Super 16 GB | Runs~90 tok/sQ4_K_M | Offload~13 tok/sQ4_K_M |
| RX 9070 XT 16 GB | Runs~86 tok/sQ4_K_M | Offload~13 tok/sQ4_K_M |
| RX 7800 XT 16 GB | Runs~83 tok/sQ4_K_M | Offload~13 tok/sQ4_K_M |
| Arc A770 16 GB | Runs~75 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RX 6800 XT 16 GB | Runs~68 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 5060 Ti 16 GB | Runs~60 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| RTX 4060 Ti 16 GB | Runs~39 tok/sQ4_K_M | Offload~7.9 tok/sQ4_K_M |
| RTX 4090 Laptop 16 GB | Runs~77 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 5090 32 GB | Runs~130 tok/sQ8_0 | Runs~170 tok/sQ6_K |
| RTX 4090 24 GB | Runs~97 tok/sQ6_K | Runs~130 tok/sQ4_K_M |
| RX 7900 XTX 24 GB | Runs~92 tok/sQ6_K | Runs~130 tok/sQ4_K_M |
| RTX 3090 / 3090 Ti 24 GB | Runs~90 tok/sQ6_K | Runs~120 tok/sQ4_K_M |
| RTX 5070 12 GB | Offload~13 tok/sQ4_K_M | Offload~13 tok/sQ4_K_M |
| RTX 4070 / 4070 Super 12 GB | Offload~11 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| Arc B580 12 GB | Offload~10 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RX 7700 XT 12 GB | Offload~9.7 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RX 6700 XT 12 GB | Offload~9.0 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| RTX 3060 12 GB | Offload~8.6 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| GTX 1080 Ti 11 GB | Offload~11 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 3080 10 GB | Offload~15 tok/sQ4_K_M | Offload~14 tok/sQ4_K_M |
| Arc A750 8 GB | Offload~12 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 5060 Ti 8 GB | Offload~11 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 5060 8 GB | Offload~11 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 3070 / 3070 Ti 8 GB | Offload~11 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 3060 Ti 8 GB | Offload~11 tok/sQ4_K_M | Offload~12 tok/sQ4_K_M |
| RTX 4060 Ti 8 GB | Offload~9.5 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RX 7600 8 GB | Offload~9.5 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 4060 8 GB | Offload~9.3 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| RTX 3050 8 GB | Offload~8.8 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| GTX 1660 / 1660 Super 6 GB | Offload~11 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| GTX 1650 4 GB | Offload~11 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 4080 Laptop 12 GB | Offload~9.7 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 4070 Laptop 8 GB | Offload~9.1 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| RTX 4060 Laptop 8 GB | Offload~9.1 tok/sQ4_K_M | Offload~10 tok/sQ4_K_M |
| RTX 3060 Laptop 6 GB | Offload~11 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
| RTX 4050 Laptop 6 GB | Offload~9.7 tok/sQ4_K_M | Offload~11 tok/sQ4_K_M |
Which should you run?
gpt-oss 20B is easier to run. gpt-oss 20B needs about 14.0 GB at Q4 with 8K context and fits fully on 16 of 40 GPUs. Qwen3 30B-A3B (MoE) needs 20.6 GB and fits on 4. On cards that run both, Qwen3 30B-A3B (MoE) is about 34% faster. Qwen3 30B-A3B (MoE) has more parameters (30.5B vs 21B), which usually means better answers if you have the memory.
Only the gpt-oss 20B fits on RTX 5080, RTX 5070 Ti, RTX 5060 Ti, RTX 4080 / 4080 Super, RTX 4070 Ti Super, RTX 4060 Ti, RTX 4090 Laptop and RX 9070 XT and 4 more.
See every result on its own page: gpt-oss 20B and Qwen3 30B-A3B (MoE).
Questions
Is gpt-oss 20B or Qwen3 30B-A3B (MoE) better?
gpt-oss 20B is easier to run. gpt-oss 20B needs about 14.0 GB at Q4 with 8K context and fits fully on 16 of 40 GPUs. Qwen3 30B-A3B (MoE) needs 20.6 GB and fits on 4. On cards that run both, Qwen3 30B-A3B (MoE) is about 34% faster. Qwen3 30B-A3B (MoE) has more parameters (30.5B vs 21B), which usually means better answers if you have the memory. The right pick depends on what you use it for: compare the strengths listed on each model's page.
Which needs less VRAM, gpt-oss 20B or Qwen3 30B-A3B (MoE)?
gpt-oss 20B needs about 14 GB at Q4 with 8K context, Qwen3 30B-A3B (MoE) about 21 GB. At Q8 it is 25 GB versus 36 GB.
Can I run gpt-oss 20B and Qwen3 30B-A3B (MoE) on a 12 GB GPU?
On an RTX 3060 12 GB, gpt-oss 20B only runs with part of the model in system RAM, at a few tokens per second and Qwen3 30B-A3B (MoE) only runs with part of the model in system RAM, at a few tokens per second.
More comparisons
Estimates from memory size and bandwidth at 8K context. Real speed varies with software, drivers and settings.