Model comparison
Wizard-Vicuna-30B-Uncensored vs Qwen3-30B-A3B-Thinking-2507
At Q4_K_M, Qwen3-30B-A3B-Thinking-2507 is the smaller download — 18,556,685,824 bytes against 19,621,139,904. At long context the gap widens: Qwen3-30B-A3B-Thinking-2507's KV cache at 32K is 16.3× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Wizard-Vicuna-30B-Uncensored | Qwen3-30B-A3B-Thinking-2507 | |
|---|---|---|
| Parameters | 32.5B | 30.5B |
| Architecture | llama | qwen3moe |
| Layers | 60 | 48 |
| Native context | 2,048 | 262,144 |
| Mixture of experts | no | yes, 128 experts |
| Quantizations published | 33 | 51 |
| Smallest quantization | 6.63 GiB | 7.05 GiB |
| Q4_K_M | 18.27 GiB | 17.28 GiB |
| Licence | other | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Wizard-Vicuna-30B-Uncensored | Qwen3-30B-A3B-Thinking-2507 | Ratio |
|---|---|---|---|
| 4,096 | 6.09 GiB | 0.38 GiB | 16.25× |
| 8,192 | 12.19 GiB | 0.75 GiB | 16.25× |
| 16,384 | 24.38 GiB | 1.50 GiB | 16.25× |
| 32,768 | 48.75 GiB | 3.00 GiB | 16.25× |
| 65,536 | 97.50 GiB | 6.00 GiB | 16.25× |
| 131,072 | 195.00 GiB | 12.00 GiB | 16.25× |