Model comparison
ERNIE-4.5-21B-A3B-Thinking vs Qwen3.6-27B
At Q4_K_M, ERNIE-4.5-21B-A3B-Thinking is the smaller download — 13,331,018,624 bytes against 16,547,397,888.
From the file· summed bytes, KV per layer
Side by side
| ERNIE-4.5-21B-A3B-Thinking | Qwen3.6-27B | |
|---|---|---|
| Parameters | 21.8B | 27.8B |
| Architecture | ernie4_5-moe | qwen35 |
| Layers | 28 | 64 |
| Native context | 131,072 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 70 | 40 |
| Smallest quantization | 4.17 GiB | 8.74 GiB |
| Q4_K_M | 12.42 GiB | 15.41 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | ERNIE-4.5-21B-A3B-Thinking | Qwen3.6-27B | Ratio |
|---|---|---|---|
| 4,096 | 0.22 GiB | 0.25 GiB | 1.14× |
| 8,192 | 0.44 GiB | 0.50 GiB | 1.14× |
| 16,384 | 0.88 GiB | 1.00 GiB | 1.14× |
| 32,768 | 1.75 GiB | 2.00 GiB | 1.14× |
| 65,536 | 3.50 GiB | 4.00 GiB | 1.14× |
| 131,072 | 7.00 GiB | 8.00 GiB | 1.14× |