Model comparison
Qwen3-VL-32B-Instruct-abliterated-v1 vs Qwen3-Coder-30B-A3B-Instruct
At Q4_K_M, Qwen3-Coder-30B-A3B-Instruct is the smaller download — 18,556,689,568 bytes against 19,762,151,360.
From the file· summed bytes, KV per layer
Side by side
| Qwen3-VL-32B-Instruct-abliterated-v1 | Qwen3-Coder-30B-A3B-Instruct | |
|---|---|---|
| Parameters | 33.4B | 30.5B |
| Architecture | qwen3vl | qwen3moe |
| Layers | — | 48 |
| Native context | — | 262,144 |
| Mixture of experts | no | yes, 128 experts |
| Quantizations published | 34 | 46 |
| Smallest quantization | 6.82 GiB | 7.46 GiB |
| Q4_K_M | 18.40 GiB | 17.28 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-VL-32B-Instruct-abliterated-v1 | Qwen3-Coder-30B-A3B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | — | 0.38 GiB | — |
| 8,192 | — | 0.75 GiB | — |
| 16,384 | — | 1.50 GiB | — |
| 32,768 | — | 3.00 GiB | — |
| 65,536 | — | 6.00 GiB | — |
| 131,072 | — | 12.00 GiB | — |