On NVIDIA GeForce RTX 4090 at Q4_K_M, 8,192 tokens
Qwen3-32B
vs Qwen3-30B-A3B
Qwen3-30B-A3B decodes 7.88× faster here.
Side by side · On NVIDIA GeForce RTX 4090 at Q4_K_M, 8,192 tokens
| Property | Qwen3-32B | Qwen3-30B-A3B |
|---|---|---|
| Runs | fits in device memory | fits in device memory |
| Decode | ~32 tok/sFaster than you read | ~251 tok/sFaster than you read |
| Memory needed | 22.7 GB of 24.0 GB | 20.2 GB of 24.0 GB |
| Spare or short | 1.29 GB spare | 3.84 GB spare |
| Weights | 19.8 GB | 18.6 GB |
| KV cache | 2.15 GB | 0.81 GB |
| Memory bandwidth | 1008 GB/s | 1008 GB/s |
Qwen3-32B
CAPACITY LIMITEDVerdictunvalidated
It fits, but only 1.29 GB remains; one more current-size context needs 2.15 GB.
Open in the calculator →Qwen3-30B-A3B
BALANCEDVerdictestimate
The configuration has memory headroom and no evaluated single hardware change improves founded decode speed by at least 20%.
Open in the calculator →Read this carefully
Both columns come from the same engine and the same published inputs, so the comparison is like for like. It is one model at one context length: change either and the answer can invert, which is what the calculator links above are for — each opens its side with the same model, device, format and context.
Decode figures are calibrated estimates from the published calibration, whose measured error is on the accuracy page.